From 3a5cb0badb106a20ef69939c90926a2944b4e2fe Mon Sep 17 00:00:00 2001 From: stevyhacker Date: Tue, 10 Nov 2026 22:24:51 +0100 Subject: [PATCH 1/2] Mac app: turn-by-turn live text, each turn about a second after its speaker MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The live transcript preview transcribes windows of 4 to 12 seconds, so a turn's last words wait up to 4 s for their window to fill, and windows cut mid-sentence. Meta's fast live text was the only way to get text right after each turn. A new setting, Turn-by-turn live text (Settings → Meetings → Processing, off by default), does it with the Transcribe model on this Mac: - LiveTurnCutter closes a turn after two quiet 256 ms chunks of speech detection, after one once the turn is 4 s long, or at 15 s at the quietest point of its last 1.5 s. - LiveMeetingTranscriber.runTurns hears both live files every 100 ms, and transcribes closed turns one at a time. A failed turn is retried and survives a pause of the preview. - While the model answers within 0.25 s and runs on this Mac, the turn still being spoken is transcribed again every 2 s and shown as a partial line. A transcription server hears each turn once, when it is finished. - deliversTurns is true in this mode, so Live Radar reads after each turn. Muse's fast live text keeps precedence and is unchanged. One read of the live file can stop short of the frames asked for; span reads now loop until they have them. The setting's two strings are in all 24 languages, and the Windows and Linux catalogs are regenerated from them. --- DEVELOPMENT.md | 1 + .../crates/lokalbot-i18n/src/catalogs.rs | 48 +++ LokalBot/CLISupport/SettingRegistry.swift | 3 + LokalBot/Engines/TranscriptionEngine.swift | 29 ++ LokalBot/Models/AppSettings.swift | 8 + LokalBot/Models/SettingDescriptor.swift | 1 + .../Resources/ar-EG.lproj/Localizable.strings | 2 + .../Resources/ar.lproj/Localizable.strings | 2 + .../Resources/bn.lproj/Localizable.strings | 2 + .../Resources/de.lproj/Localizable.strings | 2 + .../Resources/en.lproj/Localizable.strings | 2 + .../Resources/es.lproj/Localizable.strings | 2 + .../Resources/fr.lproj/Localizable.strings | 2 + .../Resources/ha.lproj/Localizable.strings | 2 + .../Resources/hi.lproj/Localizable.strings | 2 + .../Resources/id.lproj/Localizable.strings | 2 + .../Resources/it.lproj/Localizable.strings | 2 + .../Resources/ja.lproj/Localizable.strings | 2 + .../Resources/ko.lproj/Localizable.strings | 2 + .../Resources/mr.lproj/Localizable.strings | 2 + .../Resources/pcm.lproj/Localizable.strings | 2 + .../Resources/pt.lproj/Localizable.strings | 2 + .../Resources/ru.lproj/Localizable.strings | 2 + .../sr-Latn.lproj/Localizable.strings | 2 + .../Resources/sw.lproj/Localizable.strings | 2 + .../Resources/te.lproj/Localizable.strings | 2 + .../Resources/ur.lproj/Localizable.strings | 2 + .../Resources/vi.lproj/Localizable.strings | 2 + .../zh-Hans.lproj/Localizable.strings | 2 + .../zh-Hant.lproj/Localizable.strings | 2 + .../Services/LiveMeetingTranscriber.swift | 382 ++++++++++++++++-- LokalBot/Services/LiveTurnCutter.swift | 106 +++++ .../Services/MeetingRadar/MeetingRadar.swift | 11 +- LokalBot/Views/SettingsView.swift | 6 + LokalBotTests/LiveTurnCutterTests.swift | 119 ++++++ LokalBotTests/LiveTurnTextTests.swift | 340 ++++++++++++++++ PRIVACY.md | 10 +- 37 files changed, 1080 insertions(+), 32 deletions(-) create mode 100644 LokalBot/Services/LiveTurnCutter.swift create mode 100644 LokalBotTests/LiveTurnCutterTests.swift create mode 100644 LokalBotTests/LiveTurnTextTests.swift diff --git a/DEVELOPMENT.md b/DEVELOPMENT.md index f8678b67..43313b31 100644 --- a/DEVELOPMENT.md +++ b/DEVELOPMENT.md @@ -222,6 +222,7 @@ Built during the Meta Global AI Developer Hackathon on top of v0.11.4; see [HACK - **Meta connection (fast live text; the film and the replays):** `MetaModelAPI` (`LokalBot/Engines/MetaModelAPI.swift`) holds Meta's OpenAI-compatible endpoint, the standard Muse models and prices, and `MuseChatClient`, a small `/chat/completions` client with function tools, strict JSON schema, `reasoning_effort`, and token use. In the app only fast live text reaches Meta; `MuseChatClient` serves the hosted film and the replay harnesses, and its reply and message shapes are what `ThinkChat` hands the radar and Follow-through loops. Requests share the origin's `RemoteInferenceGate` and the same-origin redirect rule. The key is the Keychain item `meta-model-api-key` (`lokalbot-cli secret set meta-model-api`), falling back to Think's key when Think points at `api.meta.ai`. Sending also needs `https://api.meta.ai` in `approvedRemoteInferenceOrigins`, set from Settings → Meetings → Meta Muse. Contributor models are left out on purpose. - **Live Radar:** `MeetingRadar` reads `LiveMeetingTranscriber.lines` through `RadarLineFeed`, keeps its own copy of every line, and runs a pass at most every 20 seconds (sooner after 10 new lines). `MeetingRadarAnalyzer` builds the request: the new lines (labelled `L1`, `L2`, … for the whole recording), up to 24 earlier lines, items already found, and up to six passages from earlier meetings that `LibraryRadarEvidenceSource` finds with local keyword and semantic search. The reply must match a strict schema; each item must quote a new line (exact after normalization, or 80% of its words), and a contradiction must cite a passage (`E#`) or an earlier line (`L#`) it was shown. Everything else is dropped and counted as rejected. Results are saved to `radar.json` (`MeetingRadarRecord`) after each pass, with timing per pass (and tokens and cost where the model reports them). The pane sits between the live transcript and notes; the finished meeting's page shows a Live Radar section with Follow Through on the person's own commitments. - **Fast live text:** with `museLiveTextEnabled` (off by default) and the api.meta.ai approval, `LiveMeetingTranscriber` runs `MuseLiveTextTrack` (`LokalBot/Services/MeetingRadar/MuseLiveText.swift`) for each tee file instead of chunked local ASR. Each track opens a realtime session (`wss://api.meta.ai/v1/asr/realtime`, `ENDPOINTING` mode, 16 kHz PCM; the key goes in the handshake frame), streams the growing CAF in 100 ms frames, and turns `speechComplete` into a `Line` timed from its `speechStart`, with the in-progress turn as `partials`. `MuseLiveTextPacer` starts a session at most three seconds behind the live edge and pads a stalled track with silence, because the server closes streams that fall below real time. A session that ran 30 s or more reconnects at once (the server caps sessions at 60 minutes); one that fails sooner backs off like a local chunk. Speakers stay per track. The radar sees `deliversTurns` and passes after each turn, at most every 3 s (`MeetingRadar.Timing.turnInterval`). +- **Turn-by-turn live text:** with `turnByTurnLiveTextEnabled` (off by default; Muse's fast live text takes precedence), `LiveMeetingTranscriber.runTurns` replaces the 4 to 12 second windows. Every 100 ms it reads the whole 256 ms chunks each tee file gained (`LiveMeetingAudioPreparationWorker.readChunks`), asks `SpeechActivity.liveProbability` for each chunk's speech probability, and feeds it to that track's `LiveTurnCutter`: a turn closes after two quiet chunks, after one once it is 4 s long, or at 15 s at the quietest point of its last 1.5 s. Closed turns wait in `pendingTurns` and are transcribed one at a time with the Transcribe model, each as its own file (`prepareSpan`); a failed one is retried with the windows' backoff and survives a pause. Between turns, while the last decode took 0.25 s or less and the model is on this Mac, the open turn is transcribed again every 2 s of audio and shown as a `partials` line, as Muse's are. `deliversTurns` is true, so the radar passes after each turn. Cursors, the cutter and the speech-detection state live in the track state, so pause and retry continue where they stopped. [Benchmarks/LiveTurns](Benchmarks/LiveTurns/README.md) has the measurements and the opt-in replay tests. - **Floating panel:** `LiveRadarPanelController` (`LokalBot/Views/LiveRadarFloatingPanel.swift`) owns a borderless, non-activating `NSPanel` at `.floating` level (`canJoinAllSpaces`, `fullScreenAuxiliary`) hosting `LiveRadarPanelView`: the header with counts, `MeetingRadar.spotlightID` (the newest item, or the question that just got a screen answer) as a large `RadarCard`, the rest as rows, and the last transcript lines. `AppState` shows it when a recording starts with Live Radar on (`liveRadarEnabled` and `liveRadarPanelEnabled`) and from `startLiveRadar(for:)`, and hides it when the recording stops. `sharingType` is `.none` unless `liveRadarPanelHiddenFromCapture` is off (always `.readOnly` in the UI-test host, so the film can see it). Its top-left corner is kept in `UserDefaults` (`liveRadarPanelOrigin`). - **Screen answers:** with `museScreenAnswersEnabled` (off by default), each question a pass adds starts `MeetingRadar.answerFromScreen` beside the passes. `LibraryRadarScreenSource` runs `ActivityStore.searchOCR` (OR'd terms, captures before now, not LokalBot's windows) and decrypts each hit; `RadarScreenAnswers.crop` runs Vision `.accurate` text recognition on the decrypted image and keeps a band around the best-matching lines, as wide as the text in it (at most half the screen tall, JPEG, at most 1,600 px wide) plus that band's text in reading order. The text of up to three crops goes to the Think model (`RadarScreenAnswers.textMessages`) with a strict `{answer, image, quote}` schema; `RadarScreenAnswers.answer` keeps the reply only when the quote is in that crop's text. The hosted film and the replay send the crops themselves as `image_url` parts (`RadarScreenAnswers.messages`). The crop is saved as `radar-answers/.jpg` and the answer as `RadarItem.screenAnswer`; the request is recorded in `MeetingRadarRecord.screenAnswerRequests`. - **Live answers:** with `liveAnswersEnabled` (off by default), each question a pass adds also starts `MeetingRadar.answerLive`, which runs `LiveAnswerSession` (`LokalBot/Services/MeetingRadar/LiveAnswers.swift`) on the Think model (at most 6 model turns and 60 s). Memory comes first: `AppLiveAnswerEnvironment` puts the "Where it stands" notes of up to two tracked projects that share words with the question, active actions from the last 30 days, `PersonalContextBuilder` claims (when `askMemoryEnabled`), and work memory in the prompt as `[C1]`, `[C2]`, …. Then the model may call Ask's `search_meetings`, `get_meeting`, `get_action_items`, and `search_screen`, and `LiveAnswerTools` (`LiveAnswerTools.swift`) over `liveAnswerFolders`: `search_documents` (Spotlight with `mdfind -onlyin`, else file names ranked by the query's words, since Spotlight skips `~/Library`), `read_document` (PDFKit, `textutil`, and a small `.xlsx` reader, cut to the lines around the query), and in git repositories `search_code` (`git grep -F`), `read_file` (tracked files only), `recent_commits`, and `github` (`gh pr`/`gh issue`). Hidden paths, `node_modules`, and secret-looking names are never read, and commands time out after 8 s. Each tool result is labelled `[R#]`, and the `answer` tool must cite `C#`/`R#` ids with exact words; `LiveAnswerSession.ground` keeps a source only when its quote is in what it names (`RadarScreenAnswers.quoteIsOnScreen`) and titles it by the matching line (`path:line` with a GitHub permalink at the checked-out commit, a commit, `#123`). The answer is `RadarItem.answer` (`RadarAnswer`), the card shows `MeetingRadar.answerSteps` while it works, and the request is recorded in `MeetingRadarRecord.answerRequests`. diff --git a/LinuxWindows/crates/lokalbot-i18n/src/catalogs.rs b/LinuxWindows/crates/lokalbot-i18n/src/catalogs.rs index dc37eac7..f72e4293 100644 --- a/LinuxWindows/crates/lokalbot-i18n/src/catalogs.rs +++ b/LinuxWindows/crates/lokalbot-i18n/src/catalogs.rs @@ -2950,6 +2950,7 @@ static EN: &[(&str, &str)] = &[ ("Turn on day tracking", "Turn on day tracking"), ("Turn on microphone access in System Settings", "Turn on microphone access in System Settings"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "Turn on “Separate voices by speaker” to collect the voice samples remembering needs."), + ("Turn-by-turn live text", "Turn-by-turn live text"), ("Turning speech into a writing request", "Turning speech into a writing request"), ("Type a command or search meetings…", "Type a command or search meetings…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC."), @@ -3138,6 +3139,7 @@ static EN: &[(&str, &str)] = &[ ("Where autocomplete runs", "Where autocomplete runs"), ("Where your data lives", "Where your data lives"), ("Where your data lives and every permission LokalBot uses.", "Where your data lives and every permission LokalBot uses."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks."), ("Why does this moment matter?", "Why does this moment matter?"), ("Win", "Win"), @@ -6384,6 +6386,7 @@ static ZH_HANS: &[(&str, &str)] = &[ ("Turn on day tracking", "开启工作日追踪"), ("Turn on microphone access in System Settings", "在系统设置中开启麦克风访问权限"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "开启“区分不同说话人的声音”以采集记住所需的语音样本。"), + ("Turn-by-turn live text", "逐段发言实时转写"), ("Turning speech into a writing request", "正在将语音转为写作请求"), ("Type a command or search meetings…", "输入命令或搜索会议…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "输入回复,或将光标置于文本末尾。建议来自此 PC 上的自动完成模型。"), @@ -6572,6 +6575,7 @@ static ZH_HANS: &[(&str, &str)] = &[ ("Where autocomplete runs", "自动补全的运行位置"), ("Where your data lives", "数据存储位置"), ("Where your data lives and every permission LokalBot uses.", "你的数据存放在哪里,以及 LokalBot 使用的所有权限。"), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "实时转写预览运行期间,会在停顿处切分麦克风和通话音频,并用你的转写模型逐段转写每段发言,因此发言者停下约一秒后就会显示这段文字。实时雷达随后会在每段发言之后读取。未开启时,预览每次转写几秒钟的音频。"), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "实时转写预览运行期间,会将麦克风和通话音频流式传输到 Muse Voice Transcribe 1.0,它会在每段发言结束后约一秒返回该段的文字。未开启时,预览会在这台 Mac 上按 12 秒的片段进行转写。谁说了什么仍由你的麦克风和通话音轨判断,绝不来自 Muse。两条音轨每会议小时的费用约为 $0.36。"), ("Why does this moment matter?", "这个时刻为何重要?"), ("Win", "胜利"), @@ -9818,6 +9822,7 @@ static ZH_HANT: &[(&str, &str)] = &[ ("Turn on day tracking", "開啟每日追蹤"), ("Turn on microphone access in System Settings", "在「系統設定」中開啟麥克風取用權限"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "開啟「依講者區分聲音」,以收集記住講者所需的語音樣本。"), + ("Turn-by-turn live text", "逐段發言即時逐字稿"), ("Turning speech into a writing request", "正在將語音轉為寫作請求"), ("Type a command or search meetings…", "輸入指令或搜尋會議…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "輸入回复,或將遊標置於文字末尾。建議來自此 PC 上的自動完成模型。"), @@ -10006,6 +10011,7 @@ static ZH_HANT: &[(&str, &str)] = &[ ("Where autocomplete runs", "自動完成的執行位置"), ("Where your data lives", "資料儲存位置"), ("Where your data lives and every permission LokalBot uses.", "你的資料存放位置,以及 LokalBot 使用的所有權限。"), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "即時逐字稿預覽執行期間,會在停頓處切分麥克風和通話音訊,並用你的轉錄模型逐段轉錄每段發言,因此發言者停下約一秒後就會顯示這段文字。即時雷達隨後會在每段發言之後讀取。未開啟時,預覽每次轉錄幾秒鐘的音訊。"), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "即時逐字稿預覽執行期間,會將麥克風和通話音訊串流傳送到 Muse Voice Transcribe 1.0,它會在每段發言結束後約一秒傳回該段的文字。未開啟時,預覽會在這台 Mac 上以 12 秒為一段進行轉錄。誰說了什麼仍由你的麥克風和通話音軌判斷,絕不來自 Muse。兩條音軌每會議小時的費用約為 $0.36。"), ("Why does this moment matter?", "這個時刻為何重要?"), ("Win", "勝利"), @@ -13252,6 +13258,7 @@ static HI: &[(&str, &str)] = &[ ("Turn on day tracking", "दिन की ट्रैकिंग चालू करें"), ("Turn on microphone access in System Settings", "सिस्टम सेटिंग्स में माइक्रोफ़ोन की पहुँच चालू करें"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "याद रखने के लिए ज़रूरी आवाज़ के नमूने इकट्ठा करने हेतु “वक्ता के अनुसार आवाज़ें अलग करें” चालू करें।"), + ("Turn-by-turn live text", "बारी-दर-बारी लाइव टेक्स्ट"), ("Turning speech into a writing request", "बोलचाल को लेखन अनुरोध में बदला जा रहा है"), ("Type a command or search meetings…", "कोई कमांड लिखें या मीटिंग खोजें…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "उत्तर लिखें या कर्सर पाठ के अंत में रखें। सुझाव इस PC के Autocomplete मॉडल से आते हैं।"), @@ -13440,6 +13447,7 @@ static HI: &[(&str, &str)] = &[ ("Where autocomplete runs", "स्वतः पूर्णता कहाँ चलती है"), ("Where your data lives", "आपका डेटा कहाँ रहता है"), ("Where your data lives and every permission LokalBot uses.", "आपका डेटा कहाँ रहता है और LokalBot कौन-कौन सी अनुमतियाँ इस्तेमाल करता है।"), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "लाइव लिप्यंतरण का पूर्वावलोकन चलते समय माइक्रोफ़ोन और कॉल के ऑडियो को ठहरावों पर काटता है और हर बारी का लिप्यंतरण आपके लिप्यंतरण मॉडल से करता है, जिससे उसका टेक्स्ट वक्ता के रुकने के लगभग एक सेकंड बाद दिखता है। तब लाइव रडार हर बारी के बाद पढ़ता है। इसके बिना पूर्वावलोकन एक बार में कुछ सेकंड का लिप्यंतरण करता है।"), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "लाइव लिप्यंतरण का पूर्वावलोकन चलते समय माइक्रोफ़ोन और कॉल का ऑडियो Muse Voice Transcribe 1.0 को स्ट्रीम करता है, जो हर बारी का टेक्स्ट उसके ख़त्म होने के लगभग एक सेकंड बाद लौटाता है। इसके बिना पूर्वावलोकन इसी Mac पर 12 सेकंड के हिस्सों का लिप्यंतरण करता है। किसने क्या कहा, यह अब भी आपके माइक्रोफ़ोन और कॉल ट्रैक से तय होता है, Muse से कभी नहीं। दोनों ट्रैक के लिए मीटिंग के प्रति घंटे लगभग $0.36 लगते हैं।"), ("Why does this moment matter?", "यह क्षण क्यों महत्वपूर्ण है?"), ("Win", "Win"), @@ -16686,6 +16694,7 @@ static ES: &[(&str, &str)] = &[ ("Turn on day tracking", "Activar seguimiento del día"), ("Turn on microphone access in System Settings", "Activa el acceso al micrófono en Ajustes del Sistema"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "Activa «Separar voces por hablante» para recopilar las muestras de voz necesarias para recordar."), + ("Turn-by-turn live text", "Texto en directo intervención a intervención"), ("Turning speech into a writing request", "Convirtiendo voz en una solicitud de escritura"), ("Type a command or search meetings…", "Escribe un comando o busca reuniones…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "Escriba la respuesta o coloque el cursor al final del texto. Las sugerencias provienen del modelo Autocompletar de esta PC."), @@ -16874,6 +16883,7 @@ static ES: &[(&str, &str)] = &[ ("Where autocomplete runs", "Dónde se ejecuta Autocompletar"), ("Where your data lives", "Dónde están tus datos"), ("Where your data lives and every permission LokalBot uses.", "Dónde se guardan tus datos y cada permiso que usa LokalBot."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "Mientras se ejecuta la vista previa de transcripción en directo, corta el audio del micrófono y de la llamada en las pausas y transcribe cada intervención con tu modelo Transcribir, de modo que su texto aparece aproximadamente un segundo después de que la persona deja de hablar. El Radar en directo lee entonces después de cada intervención. Sin esta opción, la vista previa transcribe unos segundos cada vez."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "Mientras se ejecuta la vista previa de transcripción en directo, transmite el audio del micrófono y de la llamada a Muse Voice Transcribe 1.0, que devuelve el texto de cada intervención aproximadamente un segundo después de que termine. Sin esta opción, la vista previa transcribe fragmentos de 12 segundos en este Mac. Quién dijo qué sigue saliendo de tu micrófono y de la pista de la llamada, nunca de Muse. Cuesta unos $0.36 por hora de reunión por las dos pistas."), ("Why does this moment matter?", "¿Por qué importa este momento?"), ("Win", "Ganar"), @@ -20120,6 +20130,7 @@ static AR: &[(&str, &str)] = &[ ("Turn on day tracking", "تفعيل تتبّع اليوم"), ("Turn on microphone access in System Settings", "تفعيل الوصول إلى الميكروفون في إعدادات النظام"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "فعّل «فصل الأصوات بحسب المتحدث» لجمع العينات الصوتية اللازمة للتذكّر."), + ("Turn-by-turn live text", "النص المباشر مداخلةً بمداخلة"), ("Turning speech into a writing request", "جارٍ تحويل الكلام إلى طلب كتابة"), ("Type a command or search meetings…", "اكتب أمرًا أو ابحث في الاجتماعات…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "اكتب الرد، أو ضع المؤشر في نهاية النص. تأتي الاقتراحات من نموذج الإكمال التلقائي الموجود على هذا الكمبيوتر."), @@ -20308,6 +20319,7 @@ static AR: &[(&str, &str)] = &[ ("Where autocomplete runs", "مكان تشغيل الإكمال التلقائي"), ("Where your data lives", "مكان حفظ بياناتك"), ("Where your data lives and every permission LokalBot uses.", "مكان حفظ بياناتك وكل إذن يستخدمه LokalBot."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "أثناء تشغيل معاينة التفريغ النصي المباشر، يقطّع صوت الميكروفون والمكالمة عند فترات الصمت ويفرّغ كل مداخلة بنموذج التفريغ النصي لديك، فيظهر نصها بعد نحو ثانية من توقف المتحدث. وعندها يقرأ الرادار المباشر بعد كل مداخلة. ومن دونه، تفرّغ المعاينة بضع ثوانٍ في كل مرة."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "أثناء تشغيل معاينة التفريغ النصي المباشر، يبث صوت الميكروفون والمكالمة إلى Muse Voice Transcribe 1.0، الذي يعيد نص كل مداخلة بعد نحو ثانية من انتهائها. ومن دونه، تفرّغ المعاينة أجزاءً مدتها 12 ثانية على هذا الـ Mac. ويظل تحديد من قال ماذا مأخوذًا من الميكروفون ومسار المكالمة، لا من Muse أبدًا. تبلغ التكلفة نحو $0.36 لكل ساعة اجتماع للمسارين."), ("Why does this moment matter?", "لماذا تهم هذه اللحظة؟"), ("Win", "فوز"), @@ -23554,6 +23566,7 @@ static FR: &[(&str, &str)] = &[ ("Turn on day tracking", "Activer le suivi de la journée"), ("Turn on microphone access in System Settings", "Activer l’accès au microphone dans les Réglages Système"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "Activez « Séparer les voix par intervenant » pour recueillir les échantillons vocaux nécessaires à la mémorisation."), + ("Turn-by-turn live text", "Texte en direct par prise de parole"), ("Turning speech into a writing request", "Transformation de la parole en demande de rédaction"), ("Type a command or search meetings…", "Saisissez une commande ou recherchez des réunions…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "Tapez la réponse, ou placez le curseur à la fin du texte. Les suggestions proviennent du modèle Autocomplete sur ce PC."), @@ -23742,6 +23755,7 @@ static FR: &[(&str, &str)] = &[ ("Where autocomplete runs", "Où s’exécute la saisie prédictive"), ("Where your data lives", "Emplacement de vos données"), ("Where your data lives and every permission LokalBot uses.", "Où se trouvent vos données et chaque autorisation utilisée par LokalBot."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "Pendant l’aperçu de transcription en direct, découpe l’audio du microphone et de l’appel aux pauses et transcrit chaque prise de parole avec votre modèle de transcription, de sorte que son texte apparaît environ une seconde après que la personne a fini de parler. Le Radar en direct lit alors après chaque prise de parole. Sans cette option, l’aperçu transcrit quelques secondes à la fois."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "Pendant l’aperçu de transcription en direct, diffuse l’audio du microphone et de l’appel vers Muse Voice Transcribe 1.0, qui renvoie le texte de chaque prise de parole environ une seconde après sa fin. Sans cette option, l’aperçu transcrit des segments de 12 secondes sur ce Mac. Qui a dit quoi est toujours déterminé à partir de votre microphone et de la piste de l’appel, jamais par Muse. Coûte environ $0.36 par heure de réunion pour les deux pistes."), ("Why does this moment matter?", "Pourquoi ce moment est-il important ?"), ("Win", "Gagner"), @@ -26988,6 +27002,7 @@ static BN: &[(&str, &str)] = &[ ("Turn on day tracking", "দিনের কার্যকলাপ রাখা চালু করুন"), ("Turn on microphone access in System Settings", "সিস্টেম সেটিংসে মাইক্রোফোনের প্রবেশাধিকার চালু করুন"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "মনে রাখার জন্য প্রয়োজনীয় কণ্ঠের নমুনা সংগ্রহ করতে “বক্তা অনুযায়ী কণ্ঠ আলাদা করুন” চালু করুন।"), + ("Turn-by-turn live text", "পালা ধরে সরাসরি পাঠ্য"), ("Turning speech into a writing request", "কথা লেখার অনুরোধে রূপান্তর হচ্ছে"), ("Type a command or search meetings…", "কমান্ড লিখুন বা মিটিং খুঁজুন…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "উত্তর টাইপ করুন, অথবা পাঠ্যের শেষে কার্সার রাখুন। এই পিসিতে স্বয়ংসম্পূর্ণ মডেল থেকে পরামর্শগুলি আসে৷"), @@ -27176,6 +27191,7 @@ static BN: &[(&str, &str)] = &[ ("Where autocomplete runs", "স্বয়ংসম্পূর্ণ কোথায় চলে"), ("Where your data lives", "আপনার তথ্য যেখানে থাকে"), ("Where your data lives and every permission LokalBot uses.", "আপনার তথ্য কোথায় থাকে এবং LokalBot কোন কোন অনুমতি ব্যবহার করে।"), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "সরাসরি প্রতিলিপির প্রিভিউ চলার সময় মাইক্রোফোন ও কলের অডিও বিরতিতে ভাগ করে এবং আপনার প্রতিলিপি মডেল দিয়ে প্রতিটি কথার পালার প্রতিলিপি করে, ফলে বক্তা থামার প্রায় এক সেকেন্ড পরে তার পাঠ্য দেখা যায়। তখন সরাসরি রাডার প্রতিটি পালার পরে পড়ে। এটি ছাড়া প্রিভিউ একবারে কয়েক সেকেন্ড করে প্রতিলিপি করে।"), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "সরাসরি প্রতিলিপির প্রিভিউ চলার সময় মাইক্রোফোন ও কলের অডিও Muse Voice Transcribe 1.0-এ স্ট্রিম করে, যা প্রতিটি কথার পালার পাঠ্য সেটি শেষ হওয়ার প্রায় এক সেকেন্ড পরে ফেরত দেয়। এটি ছাড়া প্রিভিউ এই Mac-এ ১২ সেকেন্ডের খণ্ডে প্রতিলিপি করে। কে কী বলেছেন, তা এখনো আপনার মাইক্রোফোন ও কলের ট্র্যাক থেকেই আসে, কখনো Muse থেকে নয়। দুটি ট্র্যাকের জন্য মিটিংয়ের প্রতি ঘণ্টায় প্রায় $0.36 খরচ হয়।"), ("Why does this moment matter?", "এই মুহূর্ত কেন গুরুত্বপূর্ণ?"), ("Win", "জয়"), @@ -30422,6 +30438,7 @@ static PT: &[(&str, &str)] = &[ ("Turn on day tracking", "Ativar acompanhamento do dia"), ("Turn on microphone access in System Settings", "Ative o acesso ao microfone nos Ajustes do Sistema"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "Ative “Separar vozes por falante” para coletar as amostras de voz necessárias para lembrar."), + ("Turn-by-turn live text", "Texto ao vivo fala a fala"), ("Turning speech into a writing request", "Transformando fala em um pedido de escrita"), ("Type a command or search meetings…", "Digite um comando ou busque reuniões…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "Digite a resposta ou coloque o cursor no final do texto. As sugestões vêm do modelo Autocomplete neste PC."), @@ -30610,6 +30627,7 @@ static PT: &[(&str, &str)] = &[ ("Where autocomplete runs", "Onde o Autocompletar é executado"), ("Where your data lives", "Onde seus dados ficam"), ("Where your data lives and every permission LokalBot uses.", "Onde seus dados ficam e cada permissão que o LokalBot usa."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "Enquanto a prévia da transcrição ao vivo é executada, corta o áudio do microfone e da chamada nas pausas e transcreve cada fala com o seu modelo Transcrever, de modo que o texto aparece cerca de um segundo depois que a pessoa para de falar. O Radar ao vivo passa então a ler depois de cada fala. Sem isso, a prévia transcreve alguns segundos por vez."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "Enquanto a prévia da transcrição ao vivo é executada, transmite o áudio do microfone e da chamada para o Muse Voice Transcribe 1.0, que retorna o texto de cada fala cerca de um segundo depois que ela termina. Sem isso, a prévia transcreve blocos de 12 segundos neste Mac. Quem disse o quê continua vindo do seu microfone e da faixa da chamada, nunca do Muse. Custa cerca de $0.36 por hora de reunião pelas duas faixas."), ("Why does this moment matter?", "Por que este momento importa?"), ("Win", "Vitória"), @@ -33856,6 +33874,7 @@ static ID: &[(&str, &str)] = &[ ("Turn on day tracking", "Aktifkan pelacakan harian"), ("Turn on microphone access in System Settings", "Aktifkan akses mikrofon di Pengaturan Sistem"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "Aktifkan “Pisahkan suara per pembicara” untuk mengumpulkan sampel suara yang diperlukan agar dapat diingat."), + ("Turn-by-turn live text", "Transkrip langsung per giliran bicara"), ("Turning speech into a writing request", "Mengubah ucapan menjadi permintaan menulis"), ("Type a command or search meetings…", "Ketik perintah atau cari rapat…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "Ketik balasannya, atau letakkan kursor di akhir teks. Saran datang dari model Autocomplete pada PC ini."), @@ -34044,6 +34063,7 @@ static ID: &[(&str, &str)] = &[ ("Where autocomplete runs", "Tempat pelengkapan otomatis berjalan"), ("Where your data lives", "Tempat data Anda disimpan"), ("Where your data lives and every permission LokalBot uses.", "Lokasi data Anda dan setiap izin yang digunakan LokalBot."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "Selama pratinjau transkrip langsung berjalan, memotong audio mikrofon dan panggilan pada jeda dan mentranskripsikan setiap giliran bicara dengan model Transkripsikan Anda, sehingga teksnya muncul sekitar satu detik setelah pembicara berhenti. Radar Langsung lalu membaca setelah setiap giliran bicara. Tanpa fitur ini, pratinjau mentranskripsikan beberapa detik sekaligus."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "Selama pratinjau transkrip langsung berjalan, mengalirkan audio mikrofon dan panggilan ke Muse Voice Transcribe 1.0, yang mengembalikan teks setiap giliran bicara sekitar satu detik setelah giliran itu berakhir. Tanpa fitur ini, pratinjau mentranskripsikan potongan 12 detik di Mac ini. Siapa mengatakan apa tetap ditentukan dari mikrofon Anda dan trek panggilan, tidak pernah dari Muse. Biayanya sekitar $0.36 per jam rapat untuk kedua trek."), ("Why does this moment matter?", "Mengapa momen ini penting?"), ("Win", "Menang"), @@ -37290,6 +37310,7 @@ static UR: &[(&str, &str)] = &[ ("Turn on day tracking", "دن کی ٹریکنگ فعال کریں"), ("Turn on microphone access in System Settings", "سسٹم کی ترتیبات میں مائیکروفون تک رسائی فعال کریں"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "یاد رکھنے کے لیے درکار صوتی نمونے جمع کرنے کو «مقرر کے لحاظ سے آوازیں الگ کریں» فعال کریں۔"), + ("Turn-by-turn live text", "باری بہ باری براہ راست متن"), ("Turning speech into a writing request", "تقریر کو تحریری درخواست میں بدلا جا رہا ہے"), ("Type a command or search meetings…", "کمانڈ لکھیں یا اجلاس تلاش کریں…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "جواب میں ٹائپ کریں، یا ٹیکسٹ کے آخر میں کرسر لگائیں۔ تجاویز اس PC پر آٹو مکمل ماڈل سے آتی ہیں۔"), @@ -37478,6 +37499,7 @@ static UR: &[(&str, &str)] = &[ ("Where autocomplete runs", "خودکار تکمیل کہاں چلتی ہے"), ("Where your data lives", "آپ کا ڈیٹا کہاں رہتا ہے"), ("Where your data lives and every permission LokalBot uses.", "آپ کا ڈیٹا کہاں رہتا ہے اور LokalBot کون کون سی اجازتیں استعمال کرتا ہے۔"), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "براہ راست نقل نویسی کا پیش نظارہ چلنے کے دوران مائیکروفون اور کال کی آڈیو کو وقفوں پر کاٹتا ہے اور ہر باری کی نقل نویسی آپ کے نقل نویسی کے ماڈل سے کرتا ہے، چنانچہ اس کا متن بولنے والے کے رکنے کے تقریباً ایک سیکنڈ بعد ظاہر ہوتا ہے۔ پھر براہ راست ریڈار ہر باری کے بعد پڑھتا ہے۔ اس کے بغیر پیش نظارہ ایک وقت میں چند سیکنڈ کی نقل نویسی کرتا ہے۔"), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "براہ راست نقل نویسی کا پیش نظارہ چلنے کے دوران مائیکروفون اور کال کی آڈیو Muse Voice Transcribe 1.0 کو اسٹریم کرتا ہے، جو ہر باری کا متن اس کے ختم ہونے کے تقریباً ایک سیکنڈ بعد لوٹاتا ہے۔ اس کے بغیر پیش نظارہ اسی Mac پر 12 سیکنڈ کے حصوں کی نقل نویسی کرتا ہے۔ کس نے کیا کہا، یہ اب بھی آپ کے مائیکروفون اور کال کے ٹریک سے طے ہوتا ہے، Muse سے کبھی نہیں۔ دونوں ٹریکس کے لیے اجلاس کے فی گھنٹہ تقریباً $0.36 لاگت آتی ہے۔"), ("Why does this moment matter?", "یہ لمحہ کیوں اہم ہے؟"), ("Win", "جیت"), @@ -40724,6 +40746,7 @@ static RU: &[(&str, &str)] = &[ ("Turn on day tracking", "Включить отслеживание дня"), ("Turn on microphone access in System Settings", "Включите доступ к микрофону в Системных настройках"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "Включите «Разделять голоса по говорящим», чтобы собирать образцы голоса для запоминания."), + ("Turn-by-turn live text", "Текст в реальном времени по репликам"), ("Turning speech into a writing request", "Преобразование речи в запрос на письмо"), ("Type a command or search meetings…", "Введите команду или найдите встречу…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "Введите ответ или поставьте курсор в конец текста. Предложения поступают из модели автозаполнения на этом компьютере."), @@ -40912,6 +40935,7 @@ static RU: &[(&str, &str)] = &[ ("Where autocomplete runs", "Где работает автодополнение"), ("Where your data lives", "Где хранятся ваши данные"), ("Where your data lives and every permission LokalBot uses.", "Где хранятся ваши данные и все разрешения, которые использует LokalBot."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "Пока работает предпросмотр расшифровки, разрезает аудио микрофона и звонка на паузах и расшифровывает каждую реплику вашей моделью Расшифровать, так что её текст появляется примерно через секунду после того, как говорящий замолчал. Радар в реальном времени тогда читает после каждой реплики. Без этой функции предпросмотр расшифровывает по несколько секунд за раз."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "Пока работает предпросмотр расшифровки, передаёт аудио микрофона и звонка в Muse Voice Transcribe 1.0, который возвращает текст каждой реплики примерно через секунду после её окончания. Без этой функции предпросмотр расшифровывает 12-секундные фрагменты на этом Mac. Кто что сказал, по-прежнему определяется по вашему микрофону и дорожке звонка и никогда — моделью Muse. Стоит около $0.36 за час встречи для двух дорожек."), ("Why does this moment matter?", "Почему этот момент важен?"), ("Win", "Победа"), @@ -44158,6 +44182,7 @@ static DE: &[(&str, &str)] = &[ ("Turn on day tracking", "Tageserfassung einschalten"), ("Turn on microphone access in System Settings", "Mikrofonzugriff in den Systemeinstellungen einschalten"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "Aktiviere „Stimmen nach sprechender Person trennen“, um die zum Wiedererkennen nötigen Stimmproben zu sammeln."), + ("Turn-by-turn live text", "Live-Text Redebeitrag für Redebeitrag"), ("Turning speech into a writing request", "Sprache wird in einen Schreibauftrag umgewandelt"), ("Type a command or search meetings…", "Befehl eingeben oder Besprechungen suchen…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "Geben Sie die Antwort ein oder setzen Sie den Cursor an das Ende des Textes. Vorschläge stammen vom Autocomplete-Modell auf diesem PC."), @@ -44346,6 +44371,7 @@ static DE: &[(&str, &str)] = &[ ("Where autocomplete runs", "Wo die Autovervollständigung läuft"), ("Where your data lives", "Wo deine Daten liegen"), ("Where your data lives and every permission LokalBot uses.", "Wo deine Daten liegen und jede Berechtigung, die LokalBot nutzt."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "Schneidet während der Live-Transkriptvorschau Mikrofon- und Anrufaudio an Pausen und transkribiert jeden Redebeitrag mit deinem Transkriptionsmodell, sodass der Text etwa eine Sekunde nach dem letzten Wort erscheint. Das Live-Radar liest dann nach jedem Redebeitrag. Ohne diese Option transkribiert die Vorschau jeweils einige Sekunden am Stück."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "Streamt während der Live-Transkriptvorschau Mikrofon- und Anrufaudio an Muse Voice Transcribe 1.0, das den Text jedes Redebeitrags etwa eine Sekunde nach dessen Ende zurückgibt. Ohne diese Option transkribiert die Vorschau 12-Sekunden-Abschnitte auf diesem Mac. Wer was gesagt hat, kommt weiterhin von deinem Mikrofon und der Anrufspur, nie von Muse. Kostet etwa $0.36 pro Besprechungsstunde für beide Spuren."), ("Why does this moment matter?", "Warum ist dieser Moment wichtig?"), ("Win", "Gewinnen"), @@ -47592,6 +47618,7 @@ static JA: &[(&str, &str)] = &[ ("Turn on day tracking", "一日の追跡を有効にする"), ("Turn on microphone access in System Settings", "システム設定でマイクへのアクセスをオンにする"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "「話者ごとに声を分離」をオンにすると、記憶に必要な音声サンプルを収集できます。"), + ("Turn-by-turn live text", "発言ごとのリアルタイム文字起こし"), ("Turning speech into a writing request", "音声を文章作成依頼に変換中"), ("Type a command or search meetings…", "コマンドを入力するか会議を検索…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "返信を入力するか、テキストの末尾にカーソルを置きます。提案は、この PC のオートコンプリート モデルから提供されます。"), @@ -47780,6 +47807,7 @@ static JA: &[(&str, &str)] = &[ ("Where autocomplete runs", "自動補完の実行場所"), ("Where your data lives", "データの保存先"), ("Where your data lives and every permission LokalBot uses.", "データの保存場所と、LokalBotが使うすべての権限。"), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "リアルタイム文字起こしのプレビュー中、マイクと通話の音声を発言の切れ目で区切り、各発言を文字起こしモデルで文字起こしするため、話し終えてから約1秒でテキストが表示されます。リアルタイムレーダーは各発言のあとに読み取ります。オフの場合、プレビューは数秒ずつ文字起こしします。"), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "リアルタイム文字起こしのプレビュー中、マイクと通話の音声を Muse Voice Transcribe 1.0 にストリーミング送信し、各発言が終わってから約1秒でそのテキストが返されます。オフの場合、プレビューはこのMac上で12秒ごとに区切って文字起こしします。誰が何を話したかは引き続きマイクと通話のトラックから判断し、Museは使いません。費用は2つのトラックで会議1時間あたり約$0.36です。"), ("Why does this moment matter?", "この場面が重要な理由は?"), ("Win", "勝利"), @@ -51026,6 +51054,7 @@ static PCM: &[(&str, &str)] = &[ ("Turn on day tracking", "Turn on day tracking"), ("Turn on microphone access in System Settings", "Turn on microphone access for System Settings"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "Turn on “Separate voices by who dey talk” to collect voice samples wey remembering need."), + ("Turn-by-turn live text", "Live text turn by turn"), ("Turning speech into a writing request", "Dey turn speech into writing request"), ("Type a command or search meetings…", "Type command or search meetings…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "Type for the reply, or put cursor for the end of the text. Suggestions dey come from the Autocomplete model for this PC."), @@ -51214,6 +51243,7 @@ static PCM: &[(&str, &str)] = &[ ("Where autocomplete runs", "Where autocomplete dey run"), ("Where your data lives", "Where your data dey stay"), ("Where your data lives and every permission LokalBot uses.", "Where your data dey and every permission wey LokalBot dey use."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "While live transcript preview dey run, e dey cut microphone and call audio for where person pause, and e dey transcribe each turn with your Model for Transcribe, so the text go show about one second after the person stop to talk. Live Radar go then read after each turn. Without am, preview dey transcribe few seconds each time."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "While live transcript preview dey run, e dey stream microphone and call audio to Muse Voice Transcribe 1.0, wey dey return the text for each turn about one second after the turn end. Without am, preview dey transcribe 12-second chunks for this Mac. Who talk wetin still dey come from your microphone and the call track, never from Muse. E cost about $0.36 for each meeting hour for the two tracks."), ("Why does this moment matter?", "Why this moment matter?"), ("Win", "Win"), @@ -54460,6 +54490,7 @@ static AR_EG: &[(&str, &str)] = &[ ("Turn on day tracking", "فعّل تتبّع اليوم"), ("Turn on microphone access in System Settings", "فعّل الوصول إلى الميكروفون في إعدادات النظام"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "فعّل «افصل الأصوات حسب المتحدث» عشان تجمع عينات الصوت اللي التذكّر محتاجها."), + ("Turn-by-turn live text", "النص المباشر مداخلة بمداخلة"), ("Turning speech into a writing request", "جاري تحويل الكلام إلى طلب كتابة"), ("Type a command or search meetings…", "اكتب أمرًا أو ابحث في الاجتماعات…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "اكتب الرد، أو ضع المؤشر في نهاية النص. تأتي الاقتراحات من نموذج الإكمال التلقائي الموجود على هذا الكمبيوتر."), @@ -54648,6 +54679,7 @@ static AR_EG: &[(&str, &str)] = &[ ("Where autocomplete runs", "مكان تشغيل الإكمال التلقائي"), ("Where your data lives", "بياناتك بتتحفظ فين"), ("Where your data lives and every permission LokalBot uses.", "بياناتك متخزنة فين، وكل إذن LokalBot بيستخدمه."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "طول ما معاينة التفريغ النصي المباشر شغّالة، بيقطّع صوت الميكروفون والمكالمة عند السكتات وبيفرّغ كل مداخلة بنموذج التفريغ النصي بتاعك، فنصّها بيظهر بعد حوالي ثانية من ما المتكلم يسكت. وساعتها الرادار المباشر بيقرأ بعد كل مداخلة. من غيره، المعاينة بتفرّغ كام ثانية في المرة."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "طول ما معاينة التفريغ النصي المباشر شغّالة، بيبعت صوت الميكروفون والمكالمة أول بأول لـ Muse Voice Transcribe 1.0، اللي بيرجّع نص كل مداخلة بعد حوالي ثانية من ما تخلص. من غيره، المعاينة بتفرّغ أجزاء مدتها 12 ثانية على الـ Mac ده. مين قال إيه لسه بيتحدد من الميكروفون بتاعك ومسار المكالمة، ومش من Muse أبدًا. التكلفة حوالي $0.36 لكل ساعة اجتماع للمسارين."), ("Why does this moment matter?", "اللحظة دي مهمة ليه؟"), ("Win", "الفوز"), @@ -57894,6 +57926,7 @@ static MR: &[(&str, &str)] = &[ ("Turn on day tracking", "दिवसाचा मागोवा सुरू करा"), ("Turn on microphone access in System Settings", "सिस्टम सेटिंग्जमध्ये मायक्रोफोन प्रवेश सुरू करा"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "लक्षात ठेवण्यासाठी आवश्यक आवाजाचे नमुने मिळवण्यासाठी “वक्त्यानुसार आवाज वेगळे करा” सुरू करा."), + ("Turn-by-turn live text", "प्रत्येक वक्तव्यानंतर लाइव्ह मजकूर"), ("Turning speech into a writing request", "बोलण्याचे लेखन विनंतीत रूपांतर करत आहे"), ("Type a command or search meetings…", "कमांड टाइप करा किंवा बैठका शोधा…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "प्रत्युत्तर टाइप करा किंवा मजकुराच्या शेवटी कर्सर लावा. या PC वरील स्वयंपूर्ण मॉडेलवरून सूचना येतात."), @@ -58082,6 +58115,7 @@ static MR: &[(&str, &str)] = &[ ("Where autocomplete runs", "स्वयंपूर्णता कुठे चालते"), ("Where your data lives", "तुमचा डेटा कुठे राहतो"), ("Where your data lives and every permission LokalBot uses.", "तुमचा डेटा कुठे राहतो आणि LokalBot वापरत असलेली प्रत्येक परवानगी."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "लाइव्ह प्रतिलेखन पूर्वावलोकन चालू असताना मायक्रोफोन आणि कॉलचा ऑडिओ विरामांच्या ठिकाणी कापतो आणि प्रत्येक वक्तव्याचे प्रतिलेखन तुमच्या प्रतिलेखन मॉडेलने करतो, त्यामुळे वक्ता थांबल्यानंतर सुमारे एका सेकंदाने त्याचा मजकूर दिसतो. त्यानंतर लाइव्ह रडार प्रत्येक वक्तव्यानंतर वाचतो. याशिवाय पूर्वावलोकन एका वेळी काही सेकंदांचे प्रतिलेखन करते."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "लाइव्ह प्रतिलेखन पूर्वावलोकन चालू असताना मायक्रोफोन आणि कॉलचा ऑडिओ Muse Voice Transcribe 1.0 कडे स्ट्रीम करतो, जो प्रत्येक वक्तव्याचा मजकूर ते संपल्यानंतर सुमारे एका सेकंदाने परत देतो. याशिवाय पूर्वावलोकन या Mac वर 12 सेकंदांच्या तुकड्यांत प्रतिलेखन करते. कोण काय बोलले हे अजूनही तुमच्या मायक्रोफोन आणि कॉल ट्रॅकवरूनच ठरते, Muse कडून कधीही नाही. दोन्ही ट्रॅकसाठी बैठकीच्या प्रति तासाला सुमारे $0.36 खर्च येतो."), ("Why does this moment matter?", "हा क्षण महत्त्वाचा का आहे?"), ("Win", "विजय"), @@ -61328,6 +61362,7 @@ static VI: &[(&str, &str)] = &[ ("Turn on day tracking", "Bật theo dõi ngày"), ("Turn on microphone access in System Settings", "Bật quyền truy cập micrô trong Cài đặt hệ thống"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "Bật “Tách giọng theo người nói” để thu thập mẫu giọng cần cho việc ghi nhớ."), + ("Turn-by-turn live text", "Chép lời trực tiếp theo từng lượt nói"), ("Turning speech into a writing request", "Đang chuyển lời nói thành yêu cầu viết"), ("Type a command or search meetings…", "Nhập lệnh hoặc tìm cuộc họp…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "Nhập câu trả lời hoặc đặt con trỏ ở cuối văn bản. Các đề xuất đến từ mô hình Tự động hoàn thành trên PC này."), @@ -61516,6 +61551,7 @@ static VI: &[(&str, &str)] = &[ ("Where autocomplete runs", "Nơi tự động hoàn thành chạy"), ("Where your data lives", "Nơi lưu dữ liệu của bạn"), ("Where your data lives and every permission LokalBot uses.", "Nơi lưu dữ liệu của bạn và mọi quyền LokalBot sử dụng."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "Trong khi xem trước chép lời trực tiếp, cắt âm thanh micrô và cuộc gọi tại các quãng ngừng và chép lời từng lượt nói bằng mô hình Chép lời của bạn, nhờ đó văn bản của lượt nói xuất hiện khoảng một giây sau khi người nói dừng lại. Khi đó Radar trực tiếp đọc sau mỗi lượt nói. Nếu tắt, bản xem trước chép lời vài giây một lần."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "Trong khi xem trước chép lời trực tiếp, truyền trực tuyến âm thanh micrô và cuộc gọi đến Muse Voice Transcribe 1.0, dịch vụ này trả về văn bản của mỗi lượt nói khoảng một giây sau khi lượt đó kết thúc. Nếu tắt, bản xem trước chép lời theo từng đoạn 12 giây trên Mac này. Ai nói gì vẫn được xác định từ micrô của bạn và rãnh âm thanh cuộc gọi, không bao giờ từ Muse. Chi phí khoảng $0.36 mỗi giờ họp cho cả hai rãnh."), ("Why does this moment matter?", "Vì sao khoảnh khắc này quan trọng?"), ("Win", "Thắng"), @@ -64762,6 +64798,7 @@ static TE: &[(&str, &str)] = &[ ("Turn on day tracking", "రోజు ట్రాకింగ్ ప్రారంభించండి"), ("Turn on microphone access in System Settings", "సిస్టమ్ సెట్టింగ్‌లలో మైక్రోఫోన్ యాక్సెస్ ప్రారంభించండి"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "గుర్తుంచుకోవడానికి అవసరమైన స్వర నమూనాలను సేకరించేందుకు “వక్త ప్రకారం స్వరాలను వేరు చేయండి” ప్రారంభించండి."), + ("Turn-by-turn live text", "వంతుల వారీగా ప్రత్యక్ష వచనం"), ("Turning speech into a writing request", "మాటలను రచన అభ్యర్థనగా మారుస్తోంది"), ("Type a command or search meetings…", "కమాండ్ టైప్ చేయండి లేదా సమావేశాలను వెతకండి…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "ప్రత్యుత్తరాన్ని టైప్ చేయండి లేదా టెక్స్ట్ చివర కర్సర్ ఉంచండి. ఈ PCలో స్వీయపూర్తి మోడల్ నుండి సూచనలు వస్తాయి."), @@ -64950,6 +64987,7 @@ static TE: &[(&str, &str)] = &[ ("Where autocomplete runs", "స్వయంపూర్తి ఎక్కడ నడుస్తుంది"), ("Where your data lives", "మీ డేటా ఎక్కడ ఉంటుంది"), ("Where your data lives and every permission LokalBot uses.", "మీ డేటా ఎక్కడ ఉంటుంది, LokalBot వాడే ప్రతి అనుమతి."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "ప్రత్యక్ష ట్రాన్స్‌క్రిప్ట్ ప్రివ్యూ నడుస్తున్నప్పుడు మైక్రోఫోన్, కాల్ ఆడియోను విరామాల వద్ద విభజించి, ప్రతి వక్త వంతును మీ ట్రాన్స్‌క్రిప్షన్ మోడల్‌తో ట్రాన్స్‌క్రైబ్ చేస్తుంది; దాంతో వక్త ఆగిన సుమారు ఒక సెకను తర్వాత ఆ వంతు వచనం కనిపిస్తుంది. అప్పుడు ప్రత్యక్ష రాడార్ ప్రతి వంతు తర్వాత చదువుతుంది. ఇది లేకుండా, ప్రివ్యూ ఒకేసారి కొన్ని సెకన్ల చొప్పున ట్రాన్స్‌క్రైబ్ చేస్తుంది."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "ప్రత్యక్ష ట్రాన్స్‌క్రిప్ట్ ప్రివ్యూ నడుస్తున్నప్పుడు మైక్రోఫోన్, కాల్ ఆడియోను Muse Voice Transcribe 1.0కి స్ట్రీమ్ చేస్తుంది; ప్రతి వక్త వంతు ముగిసిన సుమారు ఒక సెకను తర్వాత అది ఆ వంతు వచనాన్ని తిరిగి ఇస్తుంది. ఇది లేకుండా, ప్రివ్యూ ఈ Macలో 12 సెకన్ల భాగాలుగా ట్రాన్స్‌క్రైబ్ చేస్తుంది. ఎవరు ఏమి అన్నారనేది ఇప్పటికీ మీ మైక్రోఫోన్, కాల్ ట్రాక్ నుంచే వస్తుంది, Muse నుంచి ఎప్పుడూ కాదు. రెండు ట్రాక్‌లకు కలిపి ఒక్కో సమావేశ గంటకు సుమారు $0.36 ఖర్చవుతుంది."), ("Why does this moment matter?", "ఈ క్షణం ఎందుకు ముఖ్యమైనది?"), ("Win", "విజయం"), @@ -68196,6 +68234,7 @@ static SW: &[(&str, &str)] = &[ ("Turn on day tracking", "Washa ufuatiliaji wa siku"), ("Turn on microphone access in System Settings", "Washa ufikiaji wa maikrofoni katika Mipangilio ya Mfumo"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "Washa “Tenganisha sauti kwa mzungumzaji” ili kukusanya sampuli za sauti zinazohitajika kwa kukumbuka."), + ("Turn-by-turn live text", "Maandishi mubashara zamu kwa zamu"), ("Turning speech into a writing request", "Inageuza usemi kuwa ombi la kuandika"), ("Type a command or search meetings…", "Andika amri au tafuta mikutano…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "Andika jibu, au weka kishale mwishoni mwa maandishi. Mapendekezo yanatoka kwa muundo wa Kukamilisha Kiotomatiki kwenye Kompyuta hii."), @@ -68384,6 +68423,7 @@ static SW: &[(&str, &str)] = &[ ("Where autocomplete runs", "Mahali ukamilishaji otomatiki unapoendeshwa"), ("Where your data lives", "Mahali data yako inapohifadhiwa"), ("Where your data lives and every permission LokalBot uses.", "Mahali data yako ilipo na kila ruhusa ambayo LokalBot hutumia."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "Wakati hakiki ya nakala ya sauti mubashara inaendelea, hukata sauti ya maikrofoni na ya simu kwenye vituo na kunukuu kila zamu kwa Modeli ya Unukuzi yako, hivyo maandishi yake huonekana takriban sekunde moja baada ya mzungumzaji kunyamaza. Kisha Rada Mubashara husoma baada ya kila zamu. Bila hiyo, hakiki hunukuu sekunde chache kwa wakati mmoja."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "Wakati hakiki ya nakala ya sauti mubashara inaendelea, hutiririsha sauti ya maikrofoni na ya simu kwenda Muse Voice Transcribe 1.0, ambayo hurudisha maandishi ya kila zamu takriban sekunde moja baada ya zamu kuisha. Bila hiyo, hakiki hunukuu vipande vya sekunde 12 kwenye Mac hii. Nani alisema nini bado hutokana na maikrofoni yako na mkondo wa sauti wa simu, kamwe si kutoka kwa Muse. Gharama ni takriban $0.36 kwa kila saa ya mkutano kwa mikondo yote miwili."), ("Why does this moment matter?", "Kwa nini tukio hili ni muhimu?"), ("Win", "Shinda"), @@ -71630,6 +71670,7 @@ static HA: &[(&str, &str)] = &[ ("Turn on day tracking", "Kunna bin ranar"), ("Turn on microphone access in System Settings", "Kunna damar makirufo a Saitunan Tsari"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "Kunna “Rarraba muryoyi bisa mai magana” don tattara samfuran murya da tunawa ke buƙata."), + ("Turn-by-turn live text", "Rubutu kai tsaye bayan kowace juyar magana"), ("Turning speech into a writing request", "Ana juya magana zuwa buƙatar rubutu"), ("Type a command or search meetings…", "Rubuta umarni ko bincika taruka…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "Rubuta amsa, ko sanya siginan kwamfuta a ƙarshen rubutun. Shawarwari sun fito daga samfurin Autocomplete akan wannan PC."), @@ -71818,6 +71859,7 @@ static HA: &[(&str, &str)] = &[ ("Where autocomplete runs", "Inda kammalawa kai tsaye ke gudana"), ("Where your data lives", "Inda bayananka suke"), ("Where your data lives and every permission LokalBot uses.", "Inda bayananka suke da kowane izini da LokalBot ke amfani da shi."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "Yayin da samfotin rubutun magana kai tsaye ke aiki, yana yanke sautin makirufo da na kira a wuraren dakatawa, kuma yana rubuta kowace juyar magana da samfurin rubuta magana naka, don haka rubutunta yana bayyana kusan daƙiƙa ɗaya bayan mai magana ya tsaya. Sannan Radar kai tsaye yana karantawa bayan kowace juyar magana. Idan babu shi, samfoti yana rubuta daƙiƙoƙi kaɗan a lokaci guda."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "Yayin da samfotin rubutun magana kai tsaye ke aiki, yana tura sautin makirufo da na kira zuwa Muse Voice Transcribe 1.0, wanda ke mayar da rubutun kowace juyar magana kusan daƙiƙa ɗaya bayan ta ƙare. Idan babu shi, samfoti yana rubuta gutsure na daƙiƙa 12 a wannan Mac. Wanda ya faɗi me har yanzu yana fitowa daga makirufonka da hanyar sautin kira, ba daga Muse ba. Kuɗinsa kusan $0.36 ne a kowace awa ta taro don hanyoyin sauti biyun."), ("Why does this moment matter?", "Me ya sa wannan lokacin yake da muhimmanci?"), ("Win", "Nasara"), @@ -75064,6 +75106,7 @@ static KO: &[(&str, &str)] = &[ ("Turn on day tracking", "하루 활동 추적 켜기"), ("Turn on microphone access in System Settings", "시스템 설정에서 마이크 접근 허용"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "화자를 기억하는 데 필요한 음성 샘플을 수집하려면 “화자별로 음성 분리”를 켜세요."), + ("Turn-by-turn live text", "발화 단위 실시간 전사"), ("Turning speech into a writing request", "음성을 글쓰기 요청으로 변환 중"), ("Type a command or search meetings…", "명령을 입력하거나 회의를 검색하세요…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "답장을 입력하거나 텍스트 끝에 커서를 놓습니다. 제안 사항은 이 PC의 자동 완성 모델에서 제공됩니다."), @@ -75252,6 +75295,7 @@ static KO: &[(&str, &str)] = &[ ("Where autocomplete runs", "자동 완성 실행 위치"), ("Where your data lives", "데이터 저장 위치"), ("Where your data lives and every permission LokalBot uses.", "데이터가 저장되는 위치와 LokalBot이 사용하는 모든 권한."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "실시간 전사문 미리보기가 실행되는 동안 마이크와 통화 오디오를 말이 멈추는 지점에서 나누고 각 발화를 전사 모델로 전사하므로, 말이 끝나고 약 1초 후에 텍스트가 표시됩니다. 그러면 실시간 레이더가 각 발화 뒤에 읽습니다. 끄면 미리보기가 한 번에 몇 초씩 전사합니다."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "실시간 전사문 미리보기가 실행되는 동안 마이크와 통화 오디오를 Muse Voice Transcribe 1.0으로 스트리밍하며, 각 발화가 끝나고 약 1초 후에 해당 텍스트를 받습니다. 끄면 미리보기가 이 Mac에서 12초 단위로 전사합니다. 누가 무엇을 말했는지는 여전히 마이크와 통화 트랙에서 판단하며 Muse를 사용하지 않습니다. 두 트랙 기준으로 회의 1시간당 약 $0.36의 비용이 듭니다."), ("Why does this moment matter?", "이 순간이 중요한 이유는 무엇인가요?"), ("Win", "승리"), @@ -78498,6 +78542,7 @@ static IT: &[(&str, &str)] = &[ ("Turn on day tracking", "Attiva il monitoraggio della giornata"), ("Turn on microphone access in System Settings", "Attiva l’accesso al microfono in Impostazioni di Sistema"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "Attiva “Separa le voci per interlocutore” per raccogliere i campioni vocali necessari per ricordare gli interlocutori."), + ("Turn-by-turn live text", "Testo in tempo reale intervento per intervento"), ("Turning speech into a writing request", "Trasformazione del parlato in una richiesta di scrittura"), ("Type a command or search meetings…", "Digita un comando o cerca nelle riunioni…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "Digita la risposta oppure posiziona il cursore alla fine del testo. I suggerimenti provengono dal modello di completamento automatico su questo PC."), @@ -78686,6 +78731,7 @@ static IT: &[(&str, &str)] = &[ ("Where autocomplete runs", "Dove viene eseguito il completamento automatico"), ("Where your data lives", "Dove si trovano i tuoi dati"), ("Where your data lives and every permission LokalBot uses.", "Dove si trovano i tuoi dati e ogni autorizzazione usata da LokalBot."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "Mentre è in esecuzione l’anteprima della trascrizione in tempo reale, taglia l’audio del microfono e della chiamata in corrispondenza delle pause e trascrive ogni intervento con il tuo modello Trascrivi, così il testo compare circa un secondo dopo che la persona ha smesso di parlare. Il Radar in tempo reale legge quindi dopo ogni intervento. Senza questa opzione, l’anteprima trascrive pochi secondi alla volta."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "Mentre è in esecuzione l’anteprima della trascrizione in tempo reale, trasmette l’audio del microfono e della chiamata a Muse Voice Transcribe 1.0, che restituisce il testo di ogni intervento circa un secondo dopo la sua fine. Senza questa opzione, l’anteprima trascrive blocchi di 12 secondi su questo Mac. Chi ha detto cosa viene sempre ricavato dal tuo microfono e dalla traccia della chiamata, mai da Muse. Costa circa $0.36 per ora di riunione per le due tracce."), ("Why does this moment matter?", "Perché questo momento è importante?"), ("Win", "Vinci"), @@ -81932,6 +81978,7 @@ static SR_LATN: &[(&str, &str)] = &[ ("Turn on day tracking", "Uključi praćenje dana"), ("Turn on microphone access in System Settings", "Uključi pristup mikrofonu u Sistemskim podešavanjima"), ("Turn on “Separate voices by speaker” to collect the voice samples remembering needs.", "Uključite „Razdvoji glasove po govornicima“ da biste prikupili uzorke glasa potrebne za pamćenje."), + ("Turn-by-turn live text", "Tekst uživo, repliku po repliku"), ("Turning speech into a writing request", "Pretvaranje govora u zahtev za pisanje"), ("Type a command or search meetings…", "Unesite komandu ili pretražite sastanke…"), ("Type in the reply, or put the cursor at the end of the text. Suggestions come from the Autocomplete model on this PC.", "Unesite odgovor ili stavite kursor na kraj teksta. Predlozi dolaze iz modela Autocomplete na ovom računaru."), @@ -82120,6 +82167,7 @@ static SR_LATN: &[(&str, &str)] = &[ ("Where autocomplete runs", "Gde radi automatsko dovršavanje"), ("Where your data lives", "Gde se nalaze vaši podaci"), ("Where your data lives and every permission LokalBot uses.", "Gde se čuvaju vaši podaci i svaka dozvola koju LokalBot koristi."), + ("While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.", "Dok radi pregled transkripta uživo, seče zvuk mikrofona i poziva na pauzama i transkribuje svaku repliku vašim modelom za Transkripciju, pa se njen tekst pojavljuje oko sekund nakon što govornik zastane. Radar uživo tada čita posle svake replike. Bez toga, pregled transkribuje po nekoliko sekundi odjednom."), ("While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks.", "Dok radi pregled transkripta uživo, prenosi zvuk mikrofona i poziva modelu Muse Voice Transcribe 1.0, koji vraća tekst svake replike oko sekund nakon što se ona završi. Bez toga, pregled na ovom Mac-u transkribuje delove od po 12 sekundi. Ko je šta rekao i dalje se određuje na osnovu vašeg mikrofona i zvučnog zapisa poziva, nikada na osnovu modela Muse. Košta oko $0.36 po satu sastanka za oba zvučna zapisa."), ("Why does this moment matter?", "Zašto je ovaj trenutak važan?"), ("Win", "Vin"), diff --git a/LokalBot/CLISupport/SettingRegistry.swift b/LokalBot/CLISupport/SettingRegistry.swift index 53fc78e1..1535eaef 100644 --- a/LokalBot/CLISupport/SettingRegistry.swift +++ b/LokalBot/CLISupport/SettingRegistry.swift @@ -387,6 +387,9 @@ enum SettingRegistry { entry("museScreenAnswersEnabled", .recording, "Answer meeting questions from your screen", .bool, .consent, aliases: "live radar screen memory answers questions screenshots think model", summary: "When Live Radar catches a question, give your Think model the text of crops of earlier screen captures that mention it, with their app, window title and time, to find the answer."), + entry("turnByTurnLiveTextEnabled", .recording, "Turn-by-turn live text", .bool, + aliases: "live transcript preview turns pauses realtime speech local radar", + summary: "While the live transcript preview runs, cut the audio at pauses and transcribe each turn with the Transcribe model, for text about a second after a speaker stops. Live Radar then reads after each turn."), entry("museLiveTextEnabled", .recording, "Fast live text with Muse Voice Transcribe", .bool, .consent, aliases: "meta muse voice transcribe realtime live transcript speech cloud audio", summary: "Stream meeting audio to Muse Voice Transcribe while the live transcript preview runs, for text within a second of each turn. Needs the api.meta.ai approval."), diff --git a/LokalBot/Engines/TranscriptionEngine.swift b/LokalBot/Engines/TranscriptionEngine.swift index 62c64ec6..95dbe6ea 100644 --- a/LokalBot/Engines/TranscriptionEngine.swift +++ b/LokalBot/Engines/TranscriptionEngine.swift @@ -824,6 +824,17 @@ actor CohereEngine: TranscriptionEngine { } } +/// The voice-activity model's memory of one live track, handed back and forth +/// through `SpeechActivity.liveProbability`. +struct LiveVoiceState: Sendable { + struct Unavailable: Error {} + + fileprivate var stream = VadStreamState.initial() + + init() {} + fileprivate init(stream: VadStreamState) { self.stream = stream } +} + /// Voice-activity gate over FluidAudio's Silero VAD. Lets the pipeline skip /// transcribing a track with no speech (e.g. your mic while muted for the whole /// call) rather than feeding silence to the ASR model, which can hallucinate @@ -876,6 +887,24 @@ actor SpeechActivity { await speechSegments(in: url)?.reduce(0.0) { $0 + $1.duration } } + /// How likely the next `liveChunkFrames` of a track being recorded are + /// speech. `state` is the model's memory of that track so far; pass back + /// the one returned. Unlike the file sweep above this throws when the + /// model is unavailable, so a live preview stops instead of treating + /// everything as speech. + func liveProbability(of chunk: [Float], + state: LiveVoiceState) async throws -> (probability: Float, state: LiveVoiceState) { + activeUses += 1 + defer { finishUse() } + try await prepare() + guard let manager else { throw LiveVoiceState.Unavailable() } + let result = try await manager.processStreamingChunk(chunk, state: state.stream) + return (result.probability, LiveVoiceState(stream: result.state)) + } + + /// Frames (16 kHz) the model judges at a time: 256 ms. + static let liveChunkFrames = VadManager.chunkSize + private nonisolated static func cacheKey(for url: URL) -> String { let attributes = try? FileManager.default.attributesOfItem(atPath: url.path) let mtime = (attributes?[.modificationDate] as? Date)?.timeIntervalSince1970 ?? -1 diff --git a/LokalBot/Models/AppSettings.swift b/LokalBot/Models/AppSettings.swift index 4a9f46a1..f950c624 100644 --- a/LokalBot/Models/AppSettings.swift +++ b/LokalBot/Models/AppSettings.swift @@ -166,6 +166,11 @@ struct AppSettings: Codable, Equatable { /// Muse Voice Transcribe for text within a second of each turn, instead /// of transcribing chunks on this Mac. Also needs the api.meta.ai approval. var museLiveTextEnabled: Bool = false + /// Off by default: the live transcript preview cuts each track at pauses + /// and transcribes one turn at a time with the Transcribe model, so a + /// turn's text follows its speaker by about a second, and Live Radar + /// reads after each turn. Muse's fast live text takes precedence. + var turnByTurnLiveTextEnabled: Bool = false /// Off by default: when Live Radar catches a question, the text of crops /// of earlier screen captures that mention it goes to the Think model for /// the answer. @@ -993,6 +998,7 @@ struct AppSettings: Codable, Equatable { case suggestActionCompletion case liveRadarEnabled case museLiveTextEnabled + case turnByTurnLiveTextEnabled case museScreenAnswersEnabled case liveRadarPanelEnabled case liveRadarPanelHiddenFromCapture @@ -1222,6 +1228,7 @@ struct AppSettings: Codable, Equatable { try c.encode(suggestActionCompletion, forKey: .suggestActionCompletion) try c.encode(liveRadarEnabled, forKey: .liveRadarEnabled) try c.encode(museLiveTextEnabled, forKey: .museLiveTextEnabled) + try c.encode(turnByTurnLiveTextEnabled, forKey: .turnByTurnLiveTextEnabled) try c.encode(museScreenAnswersEnabled, forKey: .museScreenAnswersEnabled) try c.encode(liveRadarPanelEnabled, forKey: .liveRadarPanelEnabled) try c.encode(liveRadarPanelHiddenFromCapture, forKey: .liveRadarPanelHiddenFromCapture) @@ -1430,6 +1437,7 @@ struct AppSettings: Codable, Equatable { autoTranscriptionVocabulary = decode(.autoTranscriptionVocabulary, defaults.autoTranscriptionVocabulary) suggestActionCompletion = decode(.suggestActionCompletion, defaults.suggestActionCompletion) museLiveTextEnabled = decode(.museLiveTextEnabled, defaults.museLiveTextEnabled) + turnByTurnLiveTextEnabled = decode(.turnByTurnLiveTextEnabled, defaults.turnByTurnLiveTextEnabled) museScreenAnswersEnabled = decode(.museScreenAnswersEnabled, defaults.museScreenAnswersEnabled) liveRadarPanelEnabled = decode(.liveRadarPanelEnabled, defaults.liveRadarPanelEnabled) liveRadarPanelHiddenFromCapture = decode(.liveRadarPanelHiddenFromCapture, defaults.liveRadarPanelHiddenFromCapture) diff --git a/LokalBot/Models/SettingDescriptor.swift b/LokalBot/Models/SettingDescriptor.swift index 39b01788..e42af68c 100644 --- a/LokalBot/Models/SettingDescriptor.swift +++ b/LokalBot/Models/SettingDescriptor.swift @@ -171,6 +171,7 @@ struct SettingDescriptor: Identifiable { .init(id: "settings.liveAnswerFolders", title: "Folders and repositories it may read", category: .recording, aliases: "live answers folders repositories documents code github"), .init(id: "settings.liveRadarPanelEnabled", title: "Show the floating Live Radar panel during meetings", category: .recording, aliases: "live radar floating panel overlay window always on top"), .init(id: "settings.liveRadarPanelHiddenFromCapture", title: "Hide the floating panel from screen sharing and recordings", category: .recording, aliases: "live radar panel screen share recording privacy hidden"), + .init(id: "settings.turnByTurnLiveTextEnabled", title: "Turn-by-turn live text", category: .recording, aliases: "live transcript preview turns pauses realtime speech local radar"), .init(id: "settings.museLiveTextEnabled", title: "Fast live text with Muse Voice Transcribe", category: .recording, aliases: "meta muse voice transcribe realtime live transcript speech cloud audio"), .init(id: "settings.meetingVisualContextEnabled", title: "Capture low-frequency visual context during meetings", category: .dayMemory, aliases: "meetingVisualContextEnabled"), .init(id: "settings.dayDigestAutoEnabled", title: "Generate the day digest automatically", category: .daily, aliases: "dayDigestAutoEnabled"), diff --git a/LokalBot/Resources/ar-EG.lproj/Localizable.strings b/LokalBot/Resources/ar-EG.lproj/Localizable.strings index 94e6a4f7..71e7f125 100644 --- a/LokalBot/Resources/ar-EG.lproj/Localizable.strings +++ b/LokalBot/Resources/ar-EG.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "النص المباشر السريع بيبعت صوت الاجتماع طول ما معاينة التفريغ النصي المباشر شغّالة، ومعاه اسمك من الإعدادات ← الإكمال التلقائي كتلميح لكتابته صح. الأصوات وعناوين الإيميل اللي في التقويم بتفضل على الـ Mac ده. ونفس الموافقة بتشمل التفكير لو وجّهته للخادم ده."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "النص المباشر السريع: ابعت صوت الاجتماع لـ Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "طول ما معاينة التفريغ النصي المباشر شغّالة، بيبعت صوت الميكروفون والمكالمة أول بأول لـ Muse Voice Transcribe 1.0، اللي بيرجّع نص كل مداخلة بعد حوالي ثانية من ما تخلص. من غيره، المعاينة بتفرّغ أجزاء مدتها 12 ثانية على الـ Mac ده. مين قال إيه لسه بيتحدد من الميكروفون بتاعك ومسار المكالمة، ومش من Muse أبدًا. التكلفة حوالي $0.36 لكل ساعة اجتماع للمسارين."; +"Turn-by-turn live text" = "النص المباشر مداخلة بمداخلة"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "طول ما معاينة التفريغ النصي المباشر شغّالة، بيقطّع صوت الميكروفون والمكالمة عند السكتات وبيفرّغ كل مداخلة بنموذج التفريغ النصي بتاعك، فنصّها بيظهر بعد حوالي ثانية من ما المتكلم يسكت. وساعتها الرادار المباشر بيقرأ بعد كل مداخلة. من غيره، المعاينة بتفرّغ كام ثانية في المرة."; "Answer meeting questions from your screen" = "جاوب على أسئلة الاجتماع من شاشتك"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "لما الرادار المباشر يلقط سؤال، LokalBot بيدوّر في ذاكرة الشاشة على الـ Mac ده على لقطات بتذكره، ويقص من كل لقطة السطور اللي بتذكره بس، ويدّي نموذج التفكير بتاعك نص لحد تلات أجزاء مقصوصة مع التطبيق وعنوان النافذة والوقت بتوعها. الإجابة بتظهر بس لو الاقتباس بتاعها موجود في نص الجزء المقصوص نفسه. لقطات الشاشة والبحث بيفضلوا على الـ Mac ده."; "Fast. About 3 s per radar pass." = "سريع. حوالي 3 ثواني لكل جولة رادار."; diff --git a/LokalBot/Resources/ar.lproj/Localizable.strings b/LokalBot/Resources/ar.lproj/Localizable.strings index 5b762b6d..8b81bb94 100644 --- a/LokalBot/Resources/ar.lproj/Localizable.strings +++ b/LokalBot/Resources/ar.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "يرسل النص المباشر السريع صوت الاجتماع أثناء تشغيل معاينة التفريغ النصي المباشر، مع اسمك من الإعدادات ← الإكمال التلقائي كتلميح للتهجئة. وتبقى الأصوات وعناوين البريد الإلكتروني في التقويم على هذا الـ Mac. وتشمل الموافقة نفسها التفكير إذا وجّهته إلى هذا الخادم."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "النص المباشر السريع: إرسال صوت الاجتماع إلى Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "أثناء تشغيل معاينة التفريغ النصي المباشر، يبث صوت الميكروفون والمكالمة إلى Muse Voice Transcribe 1.0، الذي يعيد نص كل مداخلة بعد نحو ثانية من انتهائها. ومن دونه، تفرّغ المعاينة أجزاءً مدتها 12 ثانية على هذا الـ Mac. ويظل تحديد من قال ماذا مأخوذًا من الميكروفون ومسار المكالمة، لا من Muse أبدًا. تبلغ التكلفة نحو $0.36 لكل ساعة اجتماع للمسارين."; +"Turn-by-turn live text" = "النص المباشر مداخلةً بمداخلة"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "أثناء تشغيل معاينة التفريغ النصي المباشر، يقطّع صوت الميكروفون والمكالمة عند فترات الصمت ويفرّغ كل مداخلة بنموذج التفريغ النصي لديك، فيظهر نصها بعد نحو ثانية من توقف المتحدث. وعندها يقرأ الرادار المباشر بعد كل مداخلة. ومن دونه، تفرّغ المعاينة بضع ثوانٍ في كل مرة."; "Answer meeting questions from your screen" = "الإجابة عن أسئلة الاجتماع من شاشتك"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "عندما يرصد الرادار المباشر سؤالًا، يبحث LokalBot في ذاكرة الشاشة على هذا الـ Mac عن لقطات تذكره، ويقتطع من كل لقطة الأسطر التي تذكره فقط، ويعطي نموذج التفكير لديك نص ما يصل إلى ثلاثة أجزاء مقتطعة مع تطبيقها وعنوان نافذتها ووقتها. لا تظهر الإجابة إلا إذا ورد اقتباسها في نص الجزء المقتطع نفسه. وتبقى لقطات الشاشة والبحث على هذا الـ Mac."; "Fast. About 3 s per radar pass." = "سريع. نحو 3 ثوانٍ لكل جولة رادار."; diff --git a/LokalBot/Resources/bn.lproj/Localizable.strings b/LokalBot/Resources/bn.lproj/Localizable.strings index de36ac38..5e4ede95 100644 --- a/LokalBot/Resources/bn.lproj/Localizable.strings +++ b/LokalBot/Resources/bn.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "দ্রুত সরাসরি পাঠ্য, সরাসরি প্রতিলিপির প্রিভিউ চলার সময় মিটিংয়ের অডিও পাঠায়, সঙ্গে বানানের ইঙ্গিত হিসেবে সেটিংস → স্বয়ংসম্পূর্ণ থেকে আপনার নামও। কণ্ঠ ও ক্যালেন্ডারের ইমেইল ঠিকানা এই Mac-এই থাকে। ভাবনাকে এই সার্ভারে নির্দেশ করলে একই অনুমোদন সেটির ক্ষেত্রেও প্রযোজ্য।"; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "দ্রুত সরাসরি পাঠ্য: মিটিংয়ের অডিও Muse Voice Transcribe 1.0-এ পাঠান"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "সরাসরি প্রতিলিপির প্রিভিউ চলার সময় মাইক্রোফোন ও কলের অডিও Muse Voice Transcribe 1.0-এ স্ট্রিম করে, যা প্রতিটি কথার পালার পাঠ্য সেটি শেষ হওয়ার প্রায় এক সেকেন্ড পরে ফেরত দেয়। এটি ছাড়া প্রিভিউ এই Mac-এ ১২ সেকেন্ডের খণ্ডে প্রতিলিপি করে। কে কী বলেছেন, তা এখনো আপনার মাইক্রোফোন ও কলের ট্র্যাক থেকেই আসে, কখনো Muse থেকে নয়। দুটি ট্র্যাকের জন্য মিটিংয়ের প্রতি ঘণ্টায় প্রায় $0.36 খরচ হয়।"; +"Turn-by-turn live text" = "পালা ধরে সরাসরি পাঠ্য"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "সরাসরি প্রতিলিপির প্রিভিউ চলার সময় মাইক্রোফোন ও কলের অডিও বিরতিতে ভাগ করে এবং আপনার প্রতিলিপি মডেল দিয়ে প্রতিটি কথার পালার প্রতিলিপি করে, ফলে বক্তা থামার প্রায় এক সেকেন্ড পরে তার পাঠ্য দেখা যায়। তখন সরাসরি রাডার প্রতিটি পালার পরে পড়ে। এটি ছাড়া প্রিভিউ একবারে কয়েক সেকেন্ড করে প্রতিলিপি করে।"; "Answer meeting questions from your screen" = "আপনার স্ক্রিন থেকে মিটিংয়ের প্রশ্নের উত্তর দিন"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "সরাসরি রাডার কোনো প্রশ্ন ধরলে LokalBot এই Mac-এ আপনার স্ক্রিন স্মৃতিতে সেই প্রশ্নের উল্লেখ থাকা ধারণ খোঁজে, প্রতিটিকে শুধু উল্লেখ থাকা লাইনগুলো পর্যন্ত কেটে নেয়, এবং সর্বোচ্চ তিনটি কাটা অংশের পাঠ্য তাদের অ্যাপ, উইন্ডোর শিরোনাম ও সময়সহ আপনার ভাবনা মডেলকে দেয়। উত্তর তখনই দেখানো হয়, যখন তার উদ্ধৃতি কাটা অংশের নিজস্ব পাঠ্যে থাকে। স্ক্রিনশট ও অনুসন্ধান এই Mac-এই থাকে।"; "Fast. About 3 s per radar pass." = "দ্রুত। প্রতি রাডার দফায় প্রায় ৩ সেকেন্ড।"; diff --git a/LokalBot/Resources/de.lproj/Localizable.strings b/LokalBot/Resources/de.lproj/Localizable.strings index c39f17d0..14fd24c4 100644 --- a/LokalBot/Resources/de.lproj/Localizable.strings +++ b/LokalBot/Resources/de.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "Schneller Live-Text sendet Besprechungsaudio, solange die Live-Transkriptvorschau läuft, zusammen mit deinem Namen aus Einstellungen → Autovervollständigung als Hinweis zur Schreibweise. Stimmen und E-Mail-Adressen aus dem Kalender bleiben auf diesem Mac. Dieselbe Genehmigung gilt für Denken, wenn du es auf diesen Server richtest."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "Schneller Live-Text: Besprechungsaudio an Muse Voice Transcribe 1.0 senden"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "Streamt während der Live-Transkriptvorschau Mikrofon- und Anrufaudio an Muse Voice Transcribe 1.0, das den Text jedes Redebeitrags etwa eine Sekunde nach dessen Ende zurückgibt. Ohne diese Option transkribiert die Vorschau 12-Sekunden-Abschnitte auf diesem Mac. Wer was gesagt hat, kommt weiterhin von deinem Mikrofon und der Anrufspur, nie von Muse. Kostet etwa $0.36 pro Besprechungsstunde für beide Spuren."; +"Turn-by-turn live text" = "Live-Text Redebeitrag für Redebeitrag"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "Schneidet während der Live-Transkriptvorschau Mikrofon- und Anrufaudio an Pausen und transkribiert jeden Redebeitrag mit deinem Transkriptionsmodell, sodass der Text etwa eine Sekunde nach dem letzten Wort erscheint. Das Live-Radar liest dann nach jedem Redebeitrag. Ohne diese Option transkribiert die Vorschau jeweils einige Sekunden am Stück."; "Answer meeting questions from your screen" = "Besprechungsfragen anhand deines Bildschirms beantworten"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "Wenn das Live-Radar eine Frage erfasst, durchsucht LokalBot dein Bildschirmgedächtnis auf diesem Mac nach Bildschirmfotos, die sie erwähnen, schneidet jedes auf die betreffenden Zeilen zu und gibt deinem Modell für Denken den Text von bis zu drei Ausschnitten mit App, Fenstertitel und Uhrzeit. Eine Antwort wird nur angezeigt, wenn ihr Zitat im Text des Ausschnitts selbst steht. Bildschirmfotos und die Suche bleiben auf diesem Mac."; "Fast. About 3 s per radar pass." = "Schnell. Etwa 3 s pro Radar-Durchlauf."; diff --git a/LokalBot/Resources/en.lproj/Localizable.strings b/LokalBot/Resources/en.lproj/Localizable.strings index d4dbeb63..dc4c4a26 100644 --- a/LokalBot/Resources/en.lproj/Localizable.strings +++ b/LokalBot/Resources/en.lproj/Localizable.strings @@ -1615,6 +1615,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "Fast live text: send meeting audio to Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks."; +"Turn-by-turn live text" = "Turn-by-turn live text"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time."; "Answer meeting questions from your screen" = "Answer meeting questions from your screen"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac."; "Fast. About 3 s per radar pass." = "Fast. About 3 s per radar pass."; diff --git a/LokalBot/Resources/es.lproj/Localizable.strings b/LokalBot/Resources/es.lproj/Localizable.strings index 114e0851..baa1cca9 100644 --- a/LokalBot/Resources/es.lproj/Localizable.strings +++ b/LokalBot/Resources/es.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "El texto en directo rápido envía el audio de la reunión mientras la vista previa de transcripción en directo está activa, con tu nombre de Ajustes → Autocompletar como indicación ortográfica. Las voces y las direcciones de correo del calendario permanecen en este Mac. La misma aprobación cubre Pensar si lo apuntas a este servidor."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "Texto en directo rápido: enviar el audio de las reuniones a Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "Mientras se ejecuta la vista previa de transcripción en directo, transmite el audio del micrófono y de la llamada a Muse Voice Transcribe 1.0, que devuelve el texto de cada intervención aproximadamente un segundo después de que termine. Sin esta opción, la vista previa transcribe fragmentos de 12 segundos en este Mac. Quién dijo qué sigue saliendo de tu micrófono y de la pista de la llamada, nunca de Muse. Cuesta unos $0.36 por hora de reunión por las dos pistas."; +"Turn-by-turn live text" = "Texto en directo intervención a intervención"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "Mientras se ejecuta la vista previa de transcripción en directo, corta el audio del micrófono y de la llamada en las pausas y transcribe cada intervención con tu modelo Transcribir, de modo que su texto aparece aproximadamente un segundo después de que la persona deja de hablar. El Radar en directo lee entonces después de cada intervención. Sin esta opción, la vista previa transcribe unos segundos cada vez."; "Answer meeting questions from your screen" = "Responder preguntas de reuniones desde tu pantalla"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "Cuando el Radar en directo detecta una pregunta, LokalBot busca en tu memoria de pantalla de este Mac las capturas que la mencionan, recorta cada una a las filas que la mencionan y le da a tu modelo de Pensar el texto de hasta tres recortes, con su app, el título de la ventana y la hora. Solo se muestra una respuesta cuando su cita aparece en el propio texto del recorte. Las capturas de pantalla y la búsqueda permanecen en este Mac."; "Fast. About 3 s per radar pass." = "Rápido. Unos 3 s por pasada del radar."; diff --git a/LokalBot/Resources/fr.lproj/Localizable.strings b/LokalBot/Resources/fr.lproj/Localizable.strings index b95b1505..34d72805 100644 --- a/LokalBot/Resources/fr.lproj/Localizable.strings +++ b/LokalBot/Resources/fr.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "Le texte en direct rapide envoie l’audio de la réunion pendant que l’aperçu de transcription en direct est actif, avec votre nom indiqué dans Réglages → Saisie prédictive comme indication d’orthographe. Les voix et les adresses e-mail du calendrier restent sur ce Mac. La même autorisation s’applique à la fonction Réfléchir si vous la faites pointer vers ce serveur."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "Texte en direct rapide : envoyer l’audio des réunions à Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "Pendant l’aperçu de transcription en direct, diffuse l’audio du microphone et de l’appel vers Muse Voice Transcribe 1.0, qui renvoie le texte de chaque prise de parole environ une seconde après sa fin. Sans cette option, l’aperçu transcrit des segments de 12 secondes sur ce Mac. Qui a dit quoi est toujours déterminé à partir de votre microphone et de la piste de l’appel, jamais par Muse. Coûte environ $0.36 par heure de réunion pour les deux pistes."; +"Turn-by-turn live text" = "Texte en direct par prise de parole"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "Pendant l’aperçu de transcription en direct, découpe l’audio du microphone et de l’appel aux pauses et transcrit chaque prise de parole avec votre modèle de transcription, de sorte que son texte apparaît environ une seconde après que la personne a fini de parler. Le Radar en direct lit alors après chaque prise de parole. Sans cette option, l’aperçu transcrit quelques secondes à la fois."; "Answer meeting questions from your screen" = "Répondre aux questions des réunions à partir de votre écran"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "Lorsque le Radar en direct repère une question, LokalBot recherche dans votre mémoire d’écran sur ce Mac les captures qui la mentionnent, recadre chacune sur les lignes concernées et transmet à votre modèle Réfléchir le texte de trois extraits au plus, avec leur app, le titre de la fenêtre et l’heure. Une réponse n’est affichée que si sa citation figure dans le texte même de l’extrait. Les captures d’écran et la recherche restent sur ce Mac."; "Fast. About 3 s per radar pass." = "Rapide. Environ 3 s par passe du radar."; diff --git a/LokalBot/Resources/ha.lproj/Localizable.strings b/LokalBot/Resources/ha.lproj/Localizable.strings index c271af48..e0c6202c 100644 --- a/LokalBot/Resources/ha.lproj/Localizable.strings +++ b/LokalBot/Resources/ha.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "Rubutu mai sauri kai tsaye yana aika sautin taro yayin da samfoti kai tsaye ke aiki, tare da sunanka daga Saituna → Kammalawa kai tsaye a matsayin alamar rubutun kalmomi. Bayanan murya da adireshin imel na kalanda suna nan a wannan Mac. Amincewar ɗaya ta shafi Tunani idan ka nuna shi ga wannan sabar."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "Rubutu mai sauri kai tsaye: aika sautin taro zuwa Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "Yayin da samfotin rubutun magana kai tsaye ke aiki, yana tura sautin makirufo da na kira zuwa Muse Voice Transcribe 1.0, wanda ke mayar da rubutun kowace juyar magana kusan daƙiƙa ɗaya bayan ta ƙare. Idan babu shi, samfoti yana rubuta gutsure na daƙiƙa 12 a wannan Mac. Wanda ya faɗi me har yanzu yana fitowa daga makirufonka da hanyar sautin kira, ba daga Muse ba. Kuɗinsa kusan $0.36 ne a kowace awa ta taro don hanyoyin sauti biyun."; +"Turn-by-turn live text" = "Rubutu kai tsaye bayan kowace juyar magana"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "Yayin da samfotin rubutun magana kai tsaye ke aiki, yana yanke sautin makirufo da na kira a wuraren dakatawa, kuma yana rubuta kowace juyar magana da samfurin rubuta magana naka, don haka rubutunta yana bayyana kusan daƙiƙa ɗaya bayan mai magana ya tsaya. Sannan Radar kai tsaye yana karantawa bayan kowace juyar magana. Idan babu shi, samfoti yana rubuta daƙiƙoƙi kaɗan a lokaci guda."; "Answer meeting questions from your screen" = "Amsa tambayoyin taro daga allonka"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "Idan Radar kai tsaye ya kama tambaya, LokalBot yana bincika ƙwaƙwalwar allonka a wannan Mac don hotunan da suka ambace ta, yana yanke kowanne zuwa layukan da suka ambace ta, kuma yana ba samfurinka na Tunani rubutun har guntu uku tare da manhajarsu, taken taga, da lokaci. Ana nuna amsa ne kawai idan abin da ta ambato yana cikin rubutun guntun da kansa. Hotunan allo da binciken suna nan a wannan Mac."; "Fast. About 3 s per radar pass." = "Mai sauri. Kusan daƙiƙa 3 a kowane zagayen radar."; diff --git a/LokalBot/Resources/hi.lproj/Localizable.strings b/LokalBot/Resources/hi.lproj/Localizable.strings index d29ad510..c5f5f937 100644 --- a/LokalBot/Resources/hi.lproj/Localizable.strings +++ b/LokalBot/Resources/hi.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "तेज़ लाइव टेक्स्ट, लाइव लिप्यंतरण का पूर्वावलोकन चलते समय मीटिंग का ऑडियो भेजता है, और साथ ही वर्तनी के संकेत के तौर पर सेटिंग्स → स्वतः पूर्णता से आपका नाम भी। आवाज़ें व कैलेंडर के ईमेल पते इसी Mac पर रहते हैं। अगर आप विचार को इस सर्वर पर लगाते हैं, तो यही स्वीकृति उस पर भी लागू होती है।"; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "तेज़ लाइव टेक्स्ट: मीटिंग का ऑडियो Muse Voice Transcribe 1.0 को भेजें"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "लाइव लिप्यंतरण का पूर्वावलोकन चलते समय माइक्रोफ़ोन और कॉल का ऑडियो Muse Voice Transcribe 1.0 को स्ट्रीम करता है, जो हर बारी का टेक्स्ट उसके ख़त्म होने के लगभग एक सेकंड बाद लौटाता है। इसके बिना पूर्वावलोकन इसी Mac पर 12 सेकंड के हिस्सों का लिप्यंतरण करता है। किसने क्या कहा, यह अब भी आपके माइक्रोफ़ोन और कॉल ट्रैक से तय होता है, Muse से कभी नहीं। दोनों ट्रैक के लिए मीटिंग के प्रति घंटे लगभग $0.36 लगते हैं।"; +"Turn-by-turn live text" = "बारी-दर-बारी लाइव टेक्स्ट"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "लाइव लिप्यंतरण का पूर्वावलोकन चलते समय माइक्रोफ़ोन और कॉल के ऑडियो को ठहरावों पर काटता है और हर बारी का लिप्यंतरण आपके लिप्यंतरण मॉडल से करता है, जिससे उसका टेक्स्ट वक्ता के रुकने के लगभग एक सेकंड बाद दिखता है। तब लाइव रडार हर बारी के बाद पढ़ता है। इसके बिना पूर्वावलोकन एक बार में कुछ सेकंड का लिप्यंतरण करता है।"; "Answer meeting questions from your screen" = "मीटिंग के प्रश्नों का उत्तर अपनी स्क्रीन से दें"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "जब लाइव रडार कोई प्रश्न पकड़ता है, तो LokalBot इसी Mac पर आपकी स्क्रीन स्मृति में ऐसे कैप्चर खोजता है जिनमें उसका ज़िक्र हो, हर एक को सिर्फ़ उन पंक्तियों तक काटता है जिनमें ज़िक्र है, और अधिकतम तीन कटे हुए हिस्सों का टेक्स्ट उनके ऐप, विंडो शीर्षक और समय के साथ आपके विचार मॉडल को देता है। उत्तर तभी दिखाया जाता है जब उसका उद्धरण कटे हुए हिस्से के अपने टेक्स्ट में हो। स्क्रीनशॉट और खोज इसी Mac पर रहते हैं।"; "Fast. About 3 s per radar pass." = "तेज़। हर रडार दौर में लगभग 3 सेकंड।"; diff --git a/LokalBot/Resources/id.lproj/Localizable.strings b/LokalBot/Resources/id.lproj/Localizable.strings index 19d001a7..0ae7c8d2 100644 --- a/LokalBot/Resources/id.lproj/Localizable.strings +++ b/LokalBot/Resources/id.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "Transkrip langsung cepat mengirim audio rapat selama pratinjau transkrip langsung berjalan, beserta nama Anda dari Pengaturan → Pelengkapan Otomatis sebagai petunjuk ejaan. Suara dan alamat email kalender tetap di Mac ini. Persetujuan yang sama berlaku untuk Pikirkan jika Anda mengarahkannya ke server ini."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "Transkrip langsung cepat: kirim audio rapat ke Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "Selama pratinjau transkrip langsung berjalan, mengalirkan audio mikrofon dan panggilan ke Muse Voice Transcribe 1.0, yang mengembalikan teks setiap giliran bicara sekitar satu detik setelah giliran itu berakhir. Tanpa fitur ini, pratinjau mentranskripsikan potongan 12 detik di Mac ini. Siapa mengatakan apa tetap ditentukan dari mikrofon Anda dan trek panggilan, tidak pernah dari Muse. Biayanya sekitar $0.36 per jam rapat untuk kedua trek."; +"Turn-by-turn live text" = "Transkrip langsung per giliran bicara"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "Selama pratinjau transkrip langsung berjalan, memotong audio mikrofon dan panggilan pada jeda dan mentranskripsikan setiap giliran bicara dengan model Transkripsikan Anda, sehingga teksnya muncul sekitar satu detik setelah pembicara berhenti. Radar Langsung lalu membaca setelah setiap giliran bicara. Tanpa fitur ini, pratinjau mentranskripsikan beberapa detik sekaligus."; "Answer meeting questions from your screen" = "Jawab pertanyaan rapat dari layar Anda"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "Saat Radar Langsung menangkap pertanyaan, LokalBot mencari tangkapan yang menyebut pertanyaan itu di memori layar Anda di Mac ini, memotong masing-masing hanya pada baris yang relevan, lalu memberikan teks hingga tiga potongan beserta aplikasi, judul jendela, dan waktunya kepada model Pikirkan Anda. Jawaban hanya ditampilkan jika kutipannya ada dalam teks potongan itu sendiri. Tangkapan layar dan pencariannya tetap di Mac ini."; "Fast. About 3 s per radar pass." = "Cepat. Sekitar 3 dtk per putaran radar."; diff --git a/LokalBot/Resources/it.lproj/Localizable.strings b/LokalBot/Resources/it.lproj/Localizable.strings index a3530e2f..2ad0e0bd 100644 --- a/LokalBot/Resources/it.lproj/Localizable.strings +++ b/LokalBot/Resources/it.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "Il testo rapido in tempo reale invia l’audio della riunione mentre l’anteprima della trascrizione in tempo reale è attiva, con il tuo nome da Impostazioni → Completamento automatico come indicazione per l’ortografia. Le voci e gli indirizzi email del calendario restano su questo Mac. La stessa approvazione vale per Pensa se lo configuri per usare questo server."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "Testo rapido in tempo reale: invia l’audio delle riunioni a Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "Mentre è in esecuzione l’anteprima della trascrizione in tempo reale, trasmette l’audio del microfono e della chiamata a Muse Voice Transcribe 1.0, che restituisce il testo di ogni intervento circa un secondo dopo la sua fine. Senza questa opzione, l’anteprima trascrive blocchi di 12 secondi su questo Mac. Chi ha detto cosa viene sempre ricavato dal tuo microfono e dalla traccia della chiamata, mai da Muse. Costa circa $0.36 per ora di riunione per le due tracce."; +"Turn-by-turn live text" = "Testo in tempo reale intervento per intervento"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "Mentre è in esecuzione l’anteprima della trascrizione in tempo reale, taglia l’audio del microfono e della chiamata in corrispondenza delle pause e trascrive ogni intervento con il tuo modello Trascrivi, così il testo compare circa un secondo dopo che la persona ha smesso di parlare. Il Radar in tempo reale legge quindi dopo ogni intervento. Senza questa opzione, l’anteprima trascrive pochi secondi alla volta."; "Answer meeting questions from your screen" = "Rispondi alle domande delle riunioni dal tuo schermo"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "Quando il Radar in tempo reale rileva una domanda, LokalBot cerca nella memoria dello schermo di questo Mac gli screenshot che la menzionano, ritaglia ciascuno sulle righe pertinenti e fornisce al tuo modello di Pensa il testo di un massimo di tre ritagli, con la relativa app, il titolo della finestra e l’ora. Una risposta viene mostrata solo quando la sua citazione compare nel testo del ritaglio stesso. Gli screenshot e la ricerca restano su questo Mac."; "Fast. About 3 s per radar pass." = "Veloce. Circa 3 s per passata del radar."; diff --git a/LokalBot/Resources/ja.lproj/Localizable.strings b/LokalBot/Resources/ja.lproj/Localizable.strings index a7df1ac0..0c49801a 100644 --- a/LokalBot/Resources/ja.lproj/Localizable.strings +++ b/LokalBot/Resources/ja.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "高速リアルタイム文字起こしは、リアルタイム文字起こしのプレビューの実行中に、会議の音声と、表記のヒントとして「設定 → 自動補完」で設定した名前を送信します。声とカレンダーのメールアドレスはこのMac内に留まります。思考機能をこのサーバーに向けた場合も、同じ承認が適用されます。"; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "高速リアルタイム文字起こし:会議の音声を Muse Voice Transcribe 1.0 に送信"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "リアルタイム文字起こしのプレビュー中、マイクと通話の音声を Muse Voice Transcribe 1.0 にストリーミング送信し、各発言が終わってから約1秒でそのテキストが返されます。オフの場合、プレビューはこのMac上で12秒ごとに区切って文字起こしします。誰が何を話したかは引き続きマイクと通話のトラックから判断し、Museは使いません。費用は2つのトラックで会議1時間あたり約$0.36です。"; +"Turn-by-turn live text" = "発言ごとのリアルタイム文字起こし"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "リアルタイム文字起こしのプレビュー中、マイクと通話の音声を発言の切れ目で区切り、各発言を文字起こしモデルで文字起こしするため、話し終えてから約1秒でテキストが表示されます。リアルタイムレーダーは各発言のあとに読み取ります。オフの場合、プレビューは数秒ずつ文字起こしします。"; "Answer meeting questions from your screen" = "会議中の質問に画面の情報から回答"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "リアルタイムレーダーが質問を検出すると、LokalBotはこのMac上の画面の記憶からその質問に触れたキャプチャを検索し、それぞれを該当する行だけに切り抜いて、最大3枚の切り抜きのテキストをアプリ名、ウィンドウタイトル、時刻とともに思考機能のモデルに渡します。回答は、その引用が切り抜き自体のテキストに含まれる場合にのみ表示されます。スクリーンショットと検索はこのMac内に留まります。"; "Fast. About 3 s per radar pass." = "高速。レーダー解析1回あたり約3秒。"; diff --git a/LokalBot/Resources/ko.lproj/Localizable.strings b/LokalBot/Resources/ko.lproj/Localizable.strings index 1ea7d58a..d4d70c4a 100644 --- a/LokalBot/Resources/ko.lproj/Localizable.strings +++ b/LokalBot/Resources/ko.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "빠른 실시간 전사는 실시간 전사문 미리보기가 실행되는 동안 회의 오디오를 보내며, 설정 → 자동 완성의 이름을 철자 힌트로 함께 보냅니다. 목소리와 캘린더 이메일 주소는 이 Mac에 머뭅니다. 사고 기능이 이 서버를 사용하도록 설정하면 같은 승인이 사고 기능에도 적용됩니다."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "빠른 실시간 전사: 회의 오디오를 Muse Voice Transcribe 1.0으로 전송"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "실시간 전사문 미리보기가 실행되는 동안 마이크와 통화 오디오를 Muse Voice Transcribe 1.0으로 스트리밍하며, 각 발화가 끝나고 약 1초 후에 해당 텍스트를 받습니다. 끄면 미리보기가 이 Mac에서 12초 단위로 전사합니다. 누가 무엇을 말했는지는 여전히 마이크와 통화 트랙에서 판단하며 Muse를 사용하지 않습니다. 두 트랙 기준으로 회의 1시간당 약 $0.36의 비용이 듭니다."; +"Turn-by-turn live text" = "발화 단위 실시간 전사"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "실시간 전사문 미리보기가 실행되는 동안 마이크와 통화 오디오를 말이 멈추는 지점에서 나누고 각 발화를 전사 모델로 전사하므로, 말이 끝나고 약 1초 후에 텍스트가 표시됩니다. 그러면 실시간 레이더가 각 발화 뒤에 읽습니다. 끄면 미리보기가 한 번에 몇 초씩 전사합니다."; "Answer meeting questions from your screen" = "화면 정보로 회의 질문에 답변"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "실시간 레이더가 질문을 포착하면 LokalBot이 이 Mac의 화면 메모리에서 질문을 언급한 캡처를 검색하고, 각 캡처를 해당 줄만 남도록 잘라 최대 3개의 잘라낸 이미지의 텍스트를 앱, 창 제목, 시간과 함께 사고 기능 모델에 전달합니다. 답변은 인용문이 잘라낸 이미지 자체의 텍스트에 있을 때만 표시됩니다. 스크린샷과 검색은 이 Mac에 머뭅니다."; "Fast. About 3 s per radar pass." = "빠름. 레이더 분석 1회당 약 3초."; diff --git a/LokalBot/Resources/mr.lproj/Localizable.strings b/LokalBot/Resources/mr.lproj/Localizable.strings index 3045dd4f..704b71cd 100644 --- a/LokalBot/Resources/mr.lproj/Localizable.strings +++ b/LokalBot/Resources/mr.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "लाइव्ह प्रतिलेखन पूर्वावलोकन चालू असताना वेगवान लाइव्ह मजकूर बैठकीचा ऑडिओ पाठवतो, आणि सोबत स्पेलिंगची सूचना म्हणून सेटिंग्ज → स्वयंपूर्णतामधील तुमचे नावही. आवाज व कॅलेंडरमधील ईमेल पत्ते या Mac वरच राहतात. तुम्ही विचार या सर्व्हरकडे वळवल्यास हीच मंजुरी त्यालाही लागू होते."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "वेगवान लाइव्ह मजकूर: बैठकीचा ऑडिओ Muse Voice Transcribe 1.0 कडे पाठवा"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "लाइव्ह प्रतिलेखन पूर्वावलोकन चालू असताना मायक्रोफोन आणि कॉलचा ऑडिओ Muse Voice Transcribe 1.0 कडे स्ट्रीम करतो, जो प्रत्येक वक्तव्याचा मजकूर ते संपल्यानंतर सुमारे एका सेकंदाने परत देतो. याशिवाय पूर्वावलोकन या Mac वर 12 सेकंदांच्या तुकड्यांत प्रतिलेखन करते. कोण काय बोलले हे अजूनही तुमच्या मायक्रोफोन आणि कॉल ट्रॅकवरूनच ठरते, Muse कडून कधीही नाही. दोन्ही ट्रॅकसाठी बैठकीच्या प्रति तासाला सुमारे $0.36 खर्च येतो."; +"Turn-by-turn live text" = "प्रत्येक वक्तव्यानंतर लाइव्ह मजकूर"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "लाइव्ह प्रतिलेखन पूर्वावलोकन चालू असताना मायक्रोफोन आणि कॉलचा ऑडिओ विरामांच्या ठिकाणी कापतो आणि प्रत्येक वक्तव्याचे प्रतिलेखन तुमच्या प्रतिलेखन मॉडेलने करतो, त्यामुळे वक्ता थांबल्यानंतर सुमारे एका सेकंदाने त्याचा मजकूर दिसतो. त्यानंतर लाइव्ह रडार प्रत्येक वक्तव्यानंतर वाचतो. याशिवाय पूर्वावलोकन एका वेळी काही सेकंदांचे प्रतिलेखन करते."; "Answer meeting questions from your screen" = "बैठकीतील प्रश्नांची उत्तरे तुमच्या स्क्रीनवरून द्या"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "लाइव्ह रडार एखादा प्रश्न पकडतो तेव्हा LokalBot या Mac वरील तुमच्या स्क्रीन स्मृतीत त्या प्रश्नाचा उल्लेख असलेले कॅप्चर शोधतो, प्रत्येक कॅप्चर फक्त उल्लेख असलेल्या ओळींपुरता कापतो, आणि जास्तीत जास्त तीन कापलेल्या भागांचा मजकूर त्यांच्या अ‍ॅप, विंडो शीर्षक आणि वेळेसह तुमच्या विचार मॉडेलला देतो. उत्तराचे अवतरण कापलेल्या भागाच्या स्वतःच्या मजकुरात असेल तेव्हाच उत्तर दाखवले जाते. स्क्रीनशॉट आणि शोध या Mac वरच राहतात."; "Fast. About 3 s per radar pass." = "वेगवान. प्रत्येक रडार फेरीला सुमारे 3 सेकंद."; diff --git a/LokalBot/Resources/pcm.lproj/Localizable.strings b/LokalBot/Resources/pcm.lproj/Localizable.strings index e2278e3d..a770823e 100644 --- a/LokalBot/Resources/pcm.lproj/Localizable.strings +++ b/LokalBot/Resources/pcm.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "Fast live text dey send meeting audio while live preview dey run, with your name from Settings → Auto-complete as spelling hint. Voice profiles and calendar email addresses dey stay for this Mac. The same approval cover Think if you point am to this server."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "Fast live text: send meeting audio to Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "While live transcript preview dey run, e dey stream microphone and call audio to Muse Voice Transcribe 1.0, wey dey return the text for each turn about one second after the turn end. Without am, preview dey transcribe 12-second chunks for this Mac. Who talk wetin still dey come from your microphone and the call track, never from Muse. E cost about $0.36 for each meeting hour for the two tracks."; +"Turn-by-turn live text" = "Live text turn by turn"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "While live transcript preview dey run, e dey cut microphone and call audio for where person pause, and e dey transcribe each turn with your Model for Transcribe, so the text go show about one second after the person stop to talk. Live Radar go then read after each turn. Without am, preview dey transcribe few seconds each time."; "Answer meeting questions from your screen" = "Answer meeting questions from your screen"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "When Live Radar catch question, LokalBot go search your screen memory for this Mac for captures wey mention am, cut each one reach the rows wey mention am, and give your Think model the text of up to three cut pieces with their app, window title, and time. E go show answer only when im quote dey inside the cut piece own text. Screenshots and the search dey stay for this Mac."; "Fast. About 3 s per radar pass." = "Fast. About 3 seconds for each radar pass."; diff --git a/LokalBot/Resources/pt.lproj/Localizable.strings b/LokalBot/Resources/pt.lproj/Localizable.strings index 5603c0e5..664b0ff4 100644 --- a/LokalBot/Resources/pt.lproj/Localizable.strings +++ b/LokalBot/Resources/pt.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "O texto ao vivo rápido envia o áudio da reunião enquanto a prévia da transcrição ao vivo está ativa, com seu nome em Ajustes → Autocompletar como dica de ortografia. As vozes e os endereços de e-mail do calendário ficam neste Mac. A mesma aprovação vale para o Pensar se você apontá-lo para este servidor."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "Texto ao vivo rápido: enviar o áudio das reuniões para o Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "Enquanto a prévia da transcrição ao vivo é executada, transmite o áudio do microfone e da chamada para o Muse Voice Transcribe 1.0, que retorna o texto de cada fala cerca de um segundo depois que ela termina. Sem isso, a prévia transcreve blocos de 12 segundos neste Mac. Quem disse o quê continua vindo do seu microfone e da faixa da chamada, nunca do Muse. Custa cerca de $0.36 por hora de reunião pelas duas faixas."; +"Turn-by-turn live text" = "Texto ao vivo fala a fala"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "Enquanto a prévia da transcrição ao vivo é executada, corta o áudio do microfone e da chamada nas pausas e transcreve cada fala com o seu modelo Transcrever, de modo que o texto aparece cerca de um segundo depois que a pessoa para de falar. O Radar ao vivo passa então a ler depois de cada fala. Sem isso, a prévia transcreve alguns segundos por vez."; "Answer meeting questions from your screen" = "Responder a perguntas de reuniões com a sua tela"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "Quando o Radar ao vivo detecta uma pergunta, o LokalBot busca na sua memória de tela deste Mac as capturas que a mencionam, recorta cada uma nas linhas que a mencionam e passa ao seu modelo do Pensar o texto de até três recortes, com o app, o título da janela e o horário. Uma resposta só é mostrada quando a citação dela está no próprio texto do recorte. As capturas de tela e a busca ficam neste Mac."; "Fast. About 3 s per radar pass." = "Rápido. Cerca de 3 s por passada do radar."; diff --git a/LokalBot/Resources/ru.lproj/Localizable.strings b/LokalBot/Resources/ru.lproj/Localizable.strings index 7aad9d42..4fe36c35 100644 --- a/LokalBot/Resources/ru.lproj/Localizable.strings +++ b/LokalBot/Resources/ru.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "Быстрый текст в реальном времени отправляет аудио встречи, пока работает предпросмотр расшифровки, а также ваше имя из раздела Настройки → Автодополнение как подсказку по написанию. Голоса и адреса электронной почты из календаря остаются на этом Mac. То же одобрение распространяется на Думать, если вы направите его на этот сервер."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "Быстрый текст в реальном времени: отправлять аудио встреч в Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "Пока работает предпросмотр расшифровки, передаёт аудио микрофона и звонка в Muse Voice Transcribe 1.0, который возвращает текст каждой реплики примерно через секунду после её окончания. Без этой функции предпросмотр расшифровывает 12-секундные фрагменты на этом Mac. Кто что сказал, по-прежнему определяется по вашему микрофону и дорожке звонка и никогда — моделью Muse. Стоит около $0.36 за час встречи для двух дорожек."; +"Turn-by-turn live text" = "Текст в реальном времени по репликам"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "Пока работает предпросмотр расшифровки, разрезает аудио микрофона и звонка на паузах и расшифровывает каждую реплику вашей моделью Расшифровать, так что её текст появляется примерно через секунду после того, как говорящий замолчал. Радар в реальном времени тогда читает после каждой реплики. Без этой функции предпросмотр расшифровывает по несколько секунд за раз."; "Answer meeting questions from your screen" = "Отвечать на вопросы встреч по экрану"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "Когда Радар в реальном времени находит вопрос, LokalBot ищет в памяти экрана на этом Mac снимки, где он упоминается, обрезает каждый до строк с упоминанием и передаёт вашей модели Думать текст до трёх фрагментов с приложением, заголовком окна и временем. Ответ показывается, только если его цитата есть в тексте самого фрагмента. Снимки экрана и поиск остаются на этом Mac."; "Fast. About 3 s per radar pass." = "Быстрая. Около 3 с на проход радара."; diff --git a/LokalBot/Resources/sr-Latn.lproj/Localizable.strings b/LokalBot/Resources/sr-Latn.lproj/Localizable.strings index 725c5a97..3eff4c97 100644 --- a/LokalBot/Resources/sr-Latn.lproj/Localizable.strings +++ b/LokalBot/Resources/sr-Latn.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "Brzi tekst uživo šalje zvuk sastanka dok radi pregled uživo, uz vaše ime iz odeljka Podešavanja → Automatsko dovršavanje kao nagoveštaj za pravopis. Glasovni profili i adrese e-pošte iz kalendara ostaju na ovom Mac-u. Isto odobrenje važi i za Razmišljanje ako ga usmerite na ovaj server."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "Brzi tekst uživo: šalji zvuk sastanka modelu Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "Dok radi pregled transkripta uživo, prenosi zvuk mikrofona i poziva modelu Muse Voice Transcribe 1.0, koji vraća tekst svake replike oko sekund nakon što se ona završi. Bez toga, pregled na ovom Mac-u transkribuje delove od po 12 sekundi. Ko je šta rekao i dalje se određuje na osnovu vašeg mikrofona i zvučnog zapisa poziva, nikada na osnovu modela Muse. Košta oko $0.36 po satu sastanka za oba zvučna zapisa."; +"Turn-by-turn live text" = "Tekst uživo, repliku po repliku"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "Dok radi pregled transkripta uživo, seče zvuk mikrofona i poziva na pauzama i transkribuje svaku repliku vašim modelom za Transkripciju, pa se njen tekst pojavljuje oko sekund nakon što govornik zastane. Radar uživo tada čita posle svake replike. Bez toga, pregled transkribuje po nekoliko sekundi odjednom."; "Answer meeting questions from your screen" = "Odgovaraj na pitanja sa sastanka na osnovu ekrana"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "Kada Radar uživo uoči pitanje, LokalBot na ovom Mac-u pretražuje vaše pamćenje ekrana tražeći snimke u kojima se ono pominje, svaki iseca na redove u kojima se pominje i vašem modelu za Razmišljanje daje tekst do tri isečka, uz njihovu aplikaciju, naslov prozora i vreme. Odgovor se prikazuje samo kada se njegov citat nalazi u samom tekstu isečka. Snimci ekrana i pretraga ostaju na ovom Mac-u."; "Fast. About 3 s per radar pass." = "Brz. Oko 3 s po prolazu Radara."; diff --git a/LokalBot/Resources/sw.lproj/Localizable.strings b/LokalBot/Resources/sw.lproj/Localizable.strings index 08ec9d55..c7ddac90 100644 --- a/LokalBot/Resources/sw.lproj/Localizable.strings +++ b/LokalBot/Resources/sw.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "Maandishi ya haraka mubashara hutuma sauti ya mkutano wakati hakiki mubashara inaendelea, pamoja na jina lako kutoka Mipangilio → Ukamilishaji otomatiki kama kidokezo cha tahajia. Wasifu wa sauti na anwani za barua pepe za kalenda hubaki kwenye Mac hii. Idhini hiyo hiyo inahusu pia Fikiri ukiielekeza kwenye seva hii."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "Maandishi ya haraka mubashara: tuma sauti ya mkutano kwa Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "Wakati hakiki ya nakala ya sauti mubashara inaendelea, hutiririsha sauti ya maikrofoni na ya simu kwenda Muse Voice Transcribe 1.0, ambayo hurudisha maandishi ya kila zamu takriban sekunde moja baada ya zamu kuisha. Bila hiyo, hakiki hunukuu vipande vya sekunde 12 kwenye Mac hii. Nani alisema nini bado hutokana na maikrofoni yako na mkondo wa sauti wa simu, kamwe si kutoka kwa Muse. Gharama ni takriban $0.36 kwa kila saa ya mkutano kwa mikondo yote miwili."; +"Turn-by-turn live text" = "Maandishi mubashara zamu kwa zamu"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "Wakati hakiki ya nakala ya sauti mubashara inaendelea, hukata sauti ya maikrofoni na ya simu kwenye vituo na kunukuu kila zamu kwa Modeli ya Unukuzi yako, hivyo maandishi yake huonekana takriban sekunde moja baada ya mzungumzaji kunyamaza. Kisha Rada Mubashara husoma baada ya kila zamu. Bila hiyo, hakiki hunukuu sekunde chache kwa wakati mmoja."; "Answer meeting questions from your screen" = "Jibu maswali ya mkutano kutoka skrini yako"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "Rada Mubashara inaponasa swali, LokalBot hutafuta kwenye kumbukumbu yako ya skrini kwenye Mac hii picha zinazolitaja, hukata kila moja hadi mistari inayolitaja, na hupa modeli yako ya Fikiri maandishi ya hadi vipande vitatu pamoja na programu, kichwa cha dirisha na wakati wake. Jibu huonyeshwa tu wakati nukuu yake iko ndani ya maandishi ya kipande chenyewe. Picha za skrini na utafutaji hubaki kwenye Mac hii."; "Fast. About 3 s per radar pass." = "Haraka. Takriban sekunde 3 kwa kila mzunguko wa rada."; diff --git a/LokalBot/Resources/te.lproj/Localizable.strings b/LokalBot/Resources/te.lproj/Localizable.strings index a82526c5..7c891e0d 100644 --- a/LokalBot/Resources/te.lproj/Localizable.strings +++ b/LokalBot/Resources/te.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "వేగవంతమైన ప్రత్యక్ష వచనం, ప్రత్యక్ష ప్రివ్యూ నడుస్తున్నంత సేపు సమావేశ ఆడియోను, అక్షరక్రమ సూచనగా సెట్టింగ్‌లు → స్వయంపూర్తి నుంచి మీ పేరును పంపుతుంది. స్వర ప్రొఫైల్‌లు, క్యాలెండర్ ఇమెయిల్ చిరునామాలు ఈ Macలోనే ఉంటాయి. మీరు ఆలోచనను ఈ సర్వర్‌కు మళ్లిస్తే అదే ఆమోదం దానికీ వర్తిస్తుంది."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "వేగవంతమైన ప్రత్యక్ష వచనం: సమావేశ ఆడియోను Muse Voice Transcribe 1.0కి పంపండి"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "ప్రత్యక్ష ట్రాన్స్‌క్రిప్ట్ ప్రివ్యూ నడుస్తున్నప్పుడు మైక్రోఫోన్, కాల్ ఆడియోను Muse Voice Transcribe 1.0కి స్ట్రీమ్ చేస్తుంది; ప్రతి వక్త వంతు ముగిసిన సుమారు ఒక సెకను తర్వాత అది ఆ వంతు వచనాన్ని తిరిగి ఇస్తుంది. ఇది లేకుండా, ప్రివ్యూ ఈ Macలో 12 సెకన్ల భాగాలుగా ట్రాన్స్‌క్రైబ్ చేస్తుంది. ఎవరు ఏమి అన్నారనేది ఇప్పటికీ మీ మైక్రోఫోన్, కాల్ ట్రాక్ నుంచే వస్తుంది, Muse నుంచి ఎప్పుడూ కాదు. రెండు ట్రాక్‌లకు కలిపి ఒక్కో సమావేశ గంటకు సుమారు $0.36 ఖర్చవుతుంది."; +"Turn-by-turn live text" = "వంతుల వారీగా ప్రత్యక్ష వచనం"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "ప్రత్యక్ష ట్రాన్స్‌క్రిప్ట్ ప్రివ్యూ నడుస్తున్నప్పుడు మైక్రోఫోన్, కాల్ ఆడియోను విరామాల వద్ద విభజించి, ప్రతి వక్త వంతును మీ ట్రాన్స్‌క్రిప్షన్ మోడల్‌తో ట్రాన్స్‌క్రైబ్ చేస్తుంది; దాంతో వక్త ఆగిన సుమారు ఒక సెకను తర్వాత ఆ వంతు వచనం కనిపిస్తుంది. అప్పుడు ప్రత్యక్ష రాడార్ ప్రతి వంతు తర్వాత చదువుతుంది. ఇది లేకుండా, ప్రివ్యూ ఒకేసారి కొన్ని సెకన్ల చొప్పున ట్రాన్స్‌క్రైబ్ చేస్తుంది."; "Answer meeting questions from your screen" = "మీ స్క్రీన్ నుంచి సమావేశ ప్రశ్నలకు సమాధానం ఇవ్వండి"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "ప్రత్యక్ష రాడార్ ఒక ప్రశ్నను గుర్తించినప్పుడు, LokalBot దాన్ని ప్రస్తావించే సంగ్రహాల కోసం ఈ Macలో మీ స్క్రీన్ జ్ఞాపకాన్ని వెతుకుతుంది, ప్రతి సంగ్రహాన్ని ప్రస్తావన ఉన్న పంక్తుల వరకు కత్తిరిస్తుంది, గరిష్ఠంగా మూడు కత్తిరింపుల వచనాన్ని వాటి యాప్, విండో శీర్షిక, సమయంతో పాటు మీ ఆలోచన మోడల్‌కు ఇస్తుంది. సమాధానంలోని ఉల్లేఖనం ఆ కత్తిరింపు సొంత వచనంలోనే ఉన్నప్పుడు మాత్రమే సమాధానం చూపబడుతుంది. స్క్రీన్‌షాట్‌లు, శోధన ఈ Macలోనే ఉంటాయి."; "Fast. About 3 s per radar pass." = "వేగవంతం. ఒక్కో రాడార్ విడతకు సుమారు 3 s."; diff --git a/LokalBot/Resources/ur.lproj/Localizable.strings b/LokalBot/Resources/ur.lproj/Localizable.strings index 5919272b..bd55608d 100644 --- a/LokalBot/Resources/ur.lproj/Localizable.strings +++ b/LokalBot/Resources/ur.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "تیز براہ راست متن، براہ راست نقل نویسی کا پیش نظارہ چلنے کے دوران اجلاس کی آڈیو بھیجتا ہے، اور ساتھ ہی ہجے کے اشارے کے طور پر ترتیبات ← خودکار تکمیل سے آپ کا نام بھی۔ آوازیں اور کیلنڈر کے ای میل پتے اسی Mac پر رہتے ہیں۔ اگر آپ سوچ کو اس سرور کی طرف موڑیں تو یہی منظوری اس پر بھی لاگو ہوتی ہے۔"; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "تیز براہ راست متن: اجلاس کی آڈیو Muse Voice Transcribe 1.0 کو بھیجیں"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "براہ راست نقل نویسی کا پیش نظارہ چلنے کے دوران مائیکروفون اور کال کی آڈیو Muse Voice Transcribe 1.0 کو اسٹریم کرتا ہے، جو ہر باری کا متن اس کے ختم ہونے کے تقریباً ایک سیکنڈ بعد لوٹاتا ہے۔ اس کے بغیر پیش نظارہ اسی Mac پر 12 سیکنڈ کے حصوں کی نقل نویسی کرتا ہے۔ کس نے کیا کہا، یہ اب بھی آپ کے مائیکروفون اور کال کے ٹریک سے طے ہوتا ہے، Muse سے کبھی نہیں۔ دونوں ٹریکس کے لیے اجلاس کے فی گھنٹہ تقریباً $0.36 لاگت آتی ہے۔"; +"Turn-by-turn live text" = "باری بہ باری براہ راست متن"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "براہ راست نقل نویسی کا پیش نظارہ چلنے کے دوران مائیکروفون اور کال کی آڈیو کو وقفوں پر کاٹتا ہے اور ہر باری کی نقل نویسی آپ کے نقل نویسی کے ماڈل سے کرتا ہے، چنانچہ اس کا متن بولنے والے کے رکنے کے تقریباً ایک سیکنڈ بعد ظاہر ہوتا ہے۔ پھر براہ راست ریڈار ہر باری کے بعد پڑھتا ہے۔ اس کے بغیر پیش نظارہ ایک وقت میں چند سیکنڈ کی نقل نویسی کرتا ہے۔"; "Answer meeting questions from your screen" = "اجلاس کے سوالات کا جواب اپنی اسکرین سے دیں"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "جب براہ راست ریڈار کوئی سوال پکڑتا ہے تو LokalBot اسی Mac پر آپ کی اسکرین کی یادداشت میں ایسے کیپچرز تلاش کرتا ہے جن میں اس کا ذکر ہو، ہر ایک کو صرف ان سطروں تک کاٹتا ہے جن میں ذکر ہے، اور زیادہ سے زیادہ تین کٹے ہوئے حصوں کا متن ان کی ایپ، ونڈو کے عنوان اور وقت کے ساتھ آپ کے سوچ ماڈل کو دیتا ہے۔ جواب صرف تب دکھایا جاتا ہے جب اس کا اقتباس کٹے ہوئے حصے کے اپنے متن میں موجود ہو۔ اسکرین شاٹس اور تلاش اسی Mac پر رہتے ہیں۔"; "Fast. About 3 s per radar pass." = "تیز۔ ہر ریڈار دور میں تقریباً 3 سیکنڈ۔"; diff --git a/LokalBot/Resources/vi.lproj/Localizable.strings b/LokalBot/Resources/vi.lproj/Localizable.strings index 280d6992..9b0cc753 100644 --- a/LokalBot/Resources/vi.lproj/Localizable.strings +++ b/LokalBot/Resources/vi.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "Chép lời trực tiếp nhanh gửi âm thanh cuộc họp trong khi xem trước chép lời trực tiếp đang chạy, kèm tên của bạn trong Cài đặt → Tự động hoàn thành làm gợi ý chính tả. Giọng nói và địa chỉ email trong lịch ở lại trên Mac này. Sự phê duyệt này cũng áp dụng cho Suy nghĩ nếu bạn trỏ Suy nghĩ đến máy chủ này."; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "Chép lời trực tiếp nhanh: gửi âm thanh cuộc họp đến Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "Trong khi xem trước chép lời trực tiếp, truyền trực tuyến âm thanh micrô và cuộc gọi đến Muse Voice Transcribe 1.0, dịch vụ này trả về văn bản của mỗi lượt nói khoảng một giây sau khi lượt đó kết thúc. Nếu tắt, bản xem trước chép lời theo từng đoạn 12 giây trên Mac này. Ai nói gì vẫn được xác định từ micrô của bạn và rãnh âm thanh cuộc gọi, không bao giờ từ Muse. Chi phí khoảng $0.36 mỗi giờ họp cho cả hai rãnh."; +"Turn-by-turn live text" = "Chép lời trực tiếp theo từng lượt nói"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "Trong khi xem trước chép lời trực tiếp, cắt âm thanh micrô và cuộc gọi tại các quãng ngừng và chép lời từng lượt nói bằng mô hình Chép lời của bạn, nhờ đó văn bản của lượt nói xuất hiện khoảng một giây sau khi người nói dừng lại. Khi đó Radar trực tiếp đọc sau mỗi lượt nói. Nếu tắt, bản xem trước chép lời vài giây một lần."; "Answer meeting questions from your screen" = "Trả lời câu hỏi trong cuộc họp từ màn hình của bạn"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "Khi Radar trực tiếp nắm bắt một câu hỏi, LokalBot tìm trong bộ nhớ màn hình trên Mac này những dữ liệu ghi lại có nhắc đến câu hỏi đó, cắt mỗi ảnh chỉ còn các dòng liên quan, rồi đưa cho mô hình Suy nghĩ của bạn văn bản của tối đa ba phần cắt kèm ứng dụng, tiêu đề cửa sổ và thời gian. Câu trả lời chỉ hiển thị khi trích dẫn của nó nằm trong chính văn bản của phần cắt. Ảnh chụp màn hình và việc tìm kiếm ở lại trên Mac này."; "Fast. About 3 s per radar pass." = "Nhanh. Khoảng 3 giây mỗi lượt phân tích radar."; diff --git a/LokalBot/Resources/zh-Hans.lproj/Localizable.strings b/LokalBot/Resources/zh-Hans.lproj/Localizable.strings index 7757379d..5bc1f074 100644 --- a/LokalBot/Resources/zh-Hans.lproj/Localizable.strings +++ b/LokalBot/Resources/zh-Hans.lproj/Localizable.strings @@ -1615,6 +1615,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "快速实时转写会在实时转写预览运行期间发送会议音频,并附上你在“设置 → 自动补全”中填写的姓名作为拼写提示。声音和日历中的电子邮件地址会留在这台 Mac 上。如果你将思考模型指向此服务器,此授权同样适用于思考模型。"; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "快速实时转写:将会议音频发送到 Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "实时转写预览运行期间,会将麦克风和通话音频流式传输到 Muse Voice Transcribe 1.0,它会在每段发言结束后约一秒返回该段的文字。未开启时,预览会在这台 Mac 上按 12 秒的片段进行转写。谁说了什么仍由你的麦克风和通话音轨判断,绝不来自 Muse。两条音轨每会议小时的费用约为 $0.36。"; +"Turn-by-turn live text" = "逐段发言实时转写"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "实时转写预览运行期间,会在停顿处切分麦克风和通话音频,并用你的转写模型逐段转写每段发言,因此发言者停下约一秒后就会显示这段文字。实时雷达随后会在每段发言之后读取。未开启时,预览每次转写几秒钟的音频。"; "Answer meeting questions from your screen" = "根据你的屏幕回答会议中的问题"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "当实时雷达捕捉到一个问题时,LokalBot 会在这台 Mac 上的屏幕记忆中搜索提及该问题的采集内容,将每张裁剪到提及该问题的那几行,并把最多三张裁剪图中的文字连同其应用、窗口标题和时间交给你的思考模型。只有当答案的引文出现在裁剪图本身的文字中时,才会显示该答案。屏幕截图和搜索过程都留在这台 Mac 上。"; "Fast. About 3 s per radar pass." = "快速。每轮雷达分析约 3 秒。"; diff --git a/LokalBot/Resources/zh-Hant.lproj/Localizable.strings b/LokalBot/Resources/zh-Hant.lproj/Localizable.strings index 2350f4f3..72645a5b 100644 --- a/LokalBot/Resources/zh-Hant.lproj/Localizable.strings +++ b/LokalBot/Resources/zh-Hant.lproj/Localizable.strings @@ -1616,6 +1616,8 @@ "Fast live text sends meeting audio while the live preview runs, with your name from Settings → Autocomplete as a spelling hint. Voices and calendar email addresses stay on this Mac. The same approval covers Think if you point it at this server." = "快速即時逐字稿會在即時逐字稿預覽執行期間傳送會議音訊,並附上你在「設定 → 自動完成」中填寫的姓名作為拼寫提示。聲音和行事曆中的電子郵件地址會留在這台 Mac 上。若你將思考模型指向此伺服器,此核准同樣適用於思考模型。"; "Fast live text: send meeting audio to Muse Voice Transcribe 1.0" = "快速即時逐字稿:將會議音訊傳送到 Muse Voice Transcribe 1.0"; "While the live transcript preview runs, streams the microphone and call audio to Muse Voice Transcribe 1.0, which returns each turn's text about a second after it ends. Without it, the preview transcribes 12-second chunks on this Mac. Who said what still comes from your microphone and the call track, never from Muse. Costs about $0.36 per meeting hour for the two tracks." = "即時逐字稿預覽執行期間,會將麥克風和通話音訊串流傳送到 Muse Voice Transcribe 1.0,它會在每段發言結束後約一秒傳回該段的文字。未開啟時,預覽會在這台 Mac 上以 12 秒為一段進行轉錄。誰說了什麼仍由你的麥克風和通話音軌判斷,絕不來自 Muse。兩條音軌每會議小時的費用約為 $0.36。"; +"Turn-by-turn live text" = "逐段發言即時逐字稿"; +"While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time." = "即時逐字稿預覽執行期間,會在停頓處切分麥克風和通話音訊,並用你的轉錄模型逐段轉錄每段發言,因此發言者停下約一秒後就會顯示這段文字。即時雷達隨後會在每段發言之後讀取。未開啟時,預覽每次轉錄幾秒鐘的音訊。"; "Answer meeting questions from your screen" = "根據你的螢幕回答會議中的問題"; "When Live Radar catches a question, LokalBot searches your screen memory on this Mac for captures that mention it, crops each to the rows that do, and gives your Think model the text of up to three crops with their app, window title, and time. An answer is shown only when its quote is in the crop's own text. Screenshots and the search stay on this Mac." = "當即時雷達捕捉到問題時,LokalBot 會在這台 Mac 上的螢幕記憶中搜尋提及該問題的擷取內容,將每張裁切到提及該問題的那幾列,並將最多三張裁切圖中的文字連同其 App、視窗標題和時間交給你的思考模型。只有在答案的引文出現在裁切圖本身的文字中時,才會顯示該答案。螢幕截圖和搜尋過程都會留在這台 Mac 上。"; "Fast. About 3 s per radar pass." = "快速。每輪雷達分析約 3 秒。"; diff --git a/LokalBot/Services/LiveMeetingTranscriber.swift b/LokalBot/Services/LiveMeetingTranscriber.swift index c468bcb2..f0eb267b 100644 --- a/LokalBot/Services/LiveMeetingTranscriber.swift +++ b/LokalBot/Services/LiveMeetingTranscriber.swift @@ -93,6 +93,68 @@ actor LiveMeetingAudioPreparationWorker { try? FileManager.default.removeItem(at: url) } + /// Whole voice-activity chunks recorded on a track since `frame`, at most + /// `maximumChunks` of them. Empty while less than one chunk is waiting or + /// the track has no file yet. + func readChunks(source: URL, from frame: Int64, chunkFrames: Int, maximumChunks: Int) throws -> [Float] { + try Task.checkCancellation() + guard frame >= 0, chunkFrames > 0, FileManager.default.fileExists(atPath: source.path) else { return [] } + let reader = try AVAudioFile(forReading: source) + try Self.requireTurnFormat(reader) + let whole = min(Int64(maximumChunks), (reader.length - frame) / Int64(chunkFrames)) + guard whole > 0 else { return [] } + // The recorder may be mid-append, so fewer frames can come back than the file reports. + let samples = try Self.read(reader, from: frame, frames: whole * Int64(chunkFrames)) + return Array(samples.prefix(samples.count / chunkFrames * chunkFrames)) + } + + /// Frames `start.. URL? { + try Task.checkCancellation() + guard start >= 0, FileManager.default.fileExists(atPath: source.path) else { return nil } + let reader = try AVAudioFile(forReading: source) + try Self.requireTurnFormat(reader) + let last = min(end, reader.length) + guard last > start else { return nil } + let samples = try Self.read(reader, from: start, frames: last - start) + guard !samples.isEmpty else { return nil } + let span = try scratchDirectory().appendingPathComponent("turn-\(UUID().uuidString).caf") + do { + try Self.writeChunk(samples, format: reader.processingFormat, to: span) + } catch { + try? FileManager.default.removeItem(at: span) + throw error + } + return span + } + + private struct UnsupportedTurnFormat: LocalizedError { + var errorDescription: String? { "The live audio file isn't 16 kHz mono." } + } + + /// Turn cutting counts frames at the rate speech detection hears them. + private static func requireTurnFormat(_ reader: AVAudioFile) throws { + guard reader.processingFormat.sampleRate == LiveMeetingTranscriber.turnSampleRate, + reader.processingFormat.channelCount == 1 else { throw UnsupportedTurnFormat() } + } + + /// One read may stop short of what was asked for, so this reads until it + /// has the frames or the file has no more. + private static func read(_ reader: AVAudioFile, from frame: Int64, frames: Int64) throws -> [Float] { + var samples: [Float] = [] + samples.reserveCapacity(Int(frames)) + reader.framePosition = frame + while Int64(samples.count) < frames, reader.framePosition < reader.length { + let wanted = AVAudioFrameCount(frames - Int64(samples.count)) + guard let buffer = AVAudioPCMBuffer(pcmFormat: reader.processingFormat, frameCapacity: wanted) else { break } + try reader.read(into: buffer, frameCount: wanted) + guard buffer.frameLength > 0, let channel = buffer.floatChannelData?[0] else { break } + samples.append(contentsOf: UnsafeBufferPointer(start: channel, count: Int(buffer.frameLength))) + } + return samples + } + /// Recent preview uses one shared time origin, even when one track is /// shorter or temporarily missing. Lagging tracks wait for that origin so /// they cannot backfill audio older than the disclosed preview start. @@ -170,17 +232,29 @@ final class LiveMeetingTranscriber: ObservableObject { enum StartPosition { case recent, beginning } - /// Where the preview's text comes from: the local model in 12-second - /// chunks, or Muse Voice Transcribe streaming turn by turn (opt-in). - enum Engine: Equatable { case local, muse } + /// Where the preview's text comes from: the Transcribe model in windows + /// of 4 to 12 seconds, the same model one turn at a time (opt-in), or + /// Muse Voice Transcribe streaming turn by turn (opt-in). + enum Engine: Equatable { case local, turns, muse } struct Timing { var idlePoll: Duration = .seconds(1) var retryBaseSeconds: Double = 1 + /// How often turn-by-turn text looks for newly recorded audio. + var turnPoll: Duration = .milliseconds(100) + /// A turn still being spoken is shown again after this much more audio. + var interimEverySeconds = 2.0 + /// It is shown only while the model answers this quickly, so a + /// finished turn never waits long behind a courtesy. + var interimBudgetSeconds = 0.25 } typealias SpeechDetector = @MainActor (URL) async throws -> Bool typealias Transcribe = @MainActor (URL, AppSettings) async throws -> Transcript + /// How likely the next 256 ms of a track are speech, with the model's + /// memory of that track handed back and forth. + typealias VoiceActivity = @Sendable ([Float], LiveVoiceState) async throws + -> (probability: Float, state: LiveVoiceState) typealias MuseTrack = (URL) -> MuseLiveTextTrack /// The key and keywords for fast live text, or nil when it can't run. typealias MuseConnection = (AppSettings) -> (key: String, keywords: [String])? @@ -204,6 +278,7 @@ final class LiveMeetingTranscriber: ObservableObject { private let audioPreparationWorker: LiveMeetingAudioPreparationWorker private let detectSpeech: SpeechDetector private let transcribe: Transcribe + private let voiceActivity: VoiceActivity private let timing: Timing private let museTrack: MuseTrack private let museConnection: MuseConnection @@ -214,14 +289,25 @@ final class LiveMeetingTranscriber: ObservableObject { private var initializedCursors = false private var resumeAfterHandoff = false private var tracks: [TrackState] + /// Turns that have closed and wait for the model, oldest first. They + /// belong to the recording like the cursors, so a pause loses none. + private var pendingTurns: [PendingTurn] = [] + /// How long the model took last time, to judge whether it is quick + /// enough to also show a turn that is still being spoken. + private var lastDecodeSeconds: Double? private static let maxLines = 400 private static let maxConsecutiveFailures = 3 + /// Speech detection hears 16 kHz, the rate both live files are written at. + nonisolated static let turnSampleRate = 16_000.0 + /// At most 8 s of one track per look, so neither track waits on a backlog. + private static let chunksPerLook = 32 nonisolated static let scratchDirectoryName = "live-previews" init(storageRoot: URL, detectSpeech: SpeechDetector? = nil, transcribe: Transcribe? = nil, + voiceActivity: VoiceActivity? = nil, timing: Timing = Timing(), museTrack: MuseTrack? = nil, museConnection: MuseConnection? = nil, @@ -245,6 +331,15 @@ final class LiveMeetingTranscriber: ObservableObject { audio: url, language: config.transcriptionLanguage.code, prompt: config.transcriptionPrompt) } + self.voiceActivity = voiceActivity ?? { chunk, state in + do { + return try await SpeechActivity.shared.liveProbability(of: chunk, state: state) + } catch is CancellationError { + throw CancellationError() + } catch { + throw PreviewError.speechDetectionUnavailable + } + } self.timing = timing } @@ -265,6 +360,7 @@ final class LiveMeetingTranscriber: ObservableObject { cancelWork() pendingFolder = folder tracks = Self.newTracks() + pendingTurns = [] initializedCursors = false hasTranscribedOnce = false resumeAfterHandoff = false @@ -304,8 +400,15 @@ final class LiveMeetingTranscriber: ObservableObject { task = Task { [weak self] in await self?.runMuse(folder: folder, session: session, key: connection.key, keywords: connection.keywords) } + } else if settings().turnByTurnLiveTextEnabled { + engine = .turns + task = Task { [weak self] in + await self?.runTurns(folder: folder, session: session, position: position) + } } else { engine = .local + // Windows carry on from the cursor, past any turn left waiting. + pendingTurns = [] task = Task { [weak self] in await self?.run(folder: folder, session: session, position: position) } @@ -326,6 +429,7 @@ final class LiveMeetingTranscriber: ObservableObject { resumeAfterHandoff = resume pendingFolder = nil tracks = Self.newTracks() + pendingTurns = [] initializedCursors = false lines = [] previewStartTime = nil @@ -392,6 +496,14 @@ final class LiveMeetingTranscriber: ObservableObject { /// session was accepted. var museCursor: Int64? var museConnectedAt: ContinuousClock.Instant? + /// Turn by turn: the cutter and speech detection's memory of this + /// track up to `processedFrames`, the last 1.5 s heard (where a + /// length cut looks for quiet), and the frame the turn in progress + /// was last shown at. + var cutter: LiveTurnCutter? + var voice = LiveVoiceState() + var tail: [Float] = [] + var interimAt: Int64 = 0 } private static func newTracks() -> [TrackState] { @@ -405,16 +517,7 @@ final class LiveMeetingTranscriber: ObservableObject { private func run(folder: URL, session: Int, position: StartPosition) async { do { - if !initializedCursors { - let cursors = try await audioPreparationWorker.initialCursors( - in: folder, fileNames: tracks.map(\.fileName), recent: position == .recent) - guard isCurrent(session) else { return } - for index in tracks.indices { tracks[index].processedFrames = cursors[index] } - // Both tee formats are fixed at 16 kHz. Actual chunk timestamps - // still use the reader's sample rate. - previewStartTime = Double(cursors.max() ?? 0) / 16_000 - initializedCursors = true - } + guard try await initializeCursors(folder: folder, session: session, position: position) else { return } while isCurrent(session) { var didWork = false for index in tracks.indices { @@ -436,18 +539,7 @@ final class LiveMeetingTranscriber: ObservableObject { } } catch { guard isCurrent(session) else { return } - tracks[index].failures += 1 - let failures = tracks[index].failures - lokalbotLog("live transcript failed track=\(tracks[index].fileName) attempt=\(failures) error=\(error.localizedDescription)") - if failures >= Self.maxConsecutiveFailures { - let message = "Live preview could not continue. Retry from this point; recording is still running." - state = .failed(message) - statusMessage = message - return - } - let delay = timing.retryBaseSeconds * pow(2, Double(failures - 1)) - tracks[index].retryAfter = ContinuousClock.now.advanced(by: .seconds(delay)) - statusMessage = "Retrying live preview… (\(failures)/\(Self.maxConsecutiveFailures))" + if fail(track: index, error: error) { return } } } // Already-buffered audio can be processed immediately. Once @@ -462,6 +554,246 @@ final class LiveMeetingTranscriber: ObservableObject { } } + /// The start choice sets each track's cursor once per recording. Returns + /// false when the session ended meanwhile. + private func initializeCursors(folder: URL, session: Int, position: StartPosition) async throws -> Bool { + guard !initializedCursors else { return true } + let cursors = try await audioPreparationWorker.initialCursors( + in: folder, fileNames: tracks.map(\.fileName), recent: position == .recent) + guard isCurrent(session) else { return false } + for index in tracks.indices { tracks[index].processedFrames = cursors[index] } + // Both tee formats are fixed at 16 kHz. Actual chunk timestamps + // still use the reader's sample rate. + previewStartTime = Double(cursors.max() ?? 0) / 16_000 + initializedCursors = true + return true + } + + /// Counts a failed step against its track and backs off. Returns true + /// when the preview has given up. + private func fail(track index: Int, error: Error) -> Bool { + tracks[index].failures += 1 + let failures = tracks[index].failures + lokalbotLog("live transcript failed track=\(tracks[index].fileName) attempt=\(failures) error=\(error.localizedDescription)") + if failures >= Self.maxConsecutiveFailures { + let message = "Live preview could not continue. Retry from this point; recording is still running." + state = .failed(message) + statusMessage = message + partials = [] + return true + } + let delay = timing.retryBaseSeconds * pow(2, Double(failures - 1)) + tracks[index].retryAfter = ContinuousClock.now.advanced(by: .seconds(delay)) + statusMessage = "Retrying live preview… (\(failures)/\(Self.maxConsecutiveFailures))" + return false + } + + // MARK: - Turn by turn on this Mac + + private struct PendingTurn { + let track: Int + let turn: LiveTurnCutter.Turn + } + + /// Hears both tracks as they are recorded, cuts them at pauses + /// (`LiveTurnCutter`), and transcribes each turn once its pause is long + /// enough: about half a second after the speaker stops, against the + /// several seconds a window takes to fill. Turns are transcribed one at a + /// time in the order they closed. In between, a model that answers + /// quickly also shows the turn still being spoken. + private func runTurns(folder: URL, session: Int, position: StartPosition) async { + do { + guard try await initializeCursors(folder: folder, session: session, position: position) else { return } + var listening = false + if !hasTranscribedOnce { statusMessage = "Preparing speech detection…" } + while isCurrent(session) { + var didWork = false + for index in tracks.indices { + guard isCurrent(session) else { return } + if let deadline = tracks[index].retryAfter, ContinuousClock.now < deadline { continue } + do { + guard try await hear(track: index, folder: folder, session: session) else { continue } + didWork = true + if !listening { + listening = true + statusMessage = nil + } + } catch { + guard isCurrent(session) else { return } + if fail(track: index, error: error) { return } + } + } + guard isCurrent(session) else { return } + dropFinishedPartials() + if let next = pendingTurns.first { + if let deadline = tracks[next.track].retryAfter, ContinuousClock.now < deadline { + // Backing off after a failure; keep hearing meanwhile. + } else { + do { + let fresh = try await decode(next, folder: folder, session: session) + guard isCurrent(session) else { return } + pendingTurns.removeFirst() + publish(fresh, for: next) + tracks[next.track].failures = 0 + tracks[next.track].retryAfter = nil + didWork = true + } catch { + guard isCurrent(session) else { return } + if fail(track: next.track, error: error) { return } + } + } + } else if try await showTurnInProgress(folder: folder, session: session) { + didWork = true + } + if !didWork { try await Task.sleep(for: timing.turnPoll) } + } + } catch { + guard isCurrent(session) else { return } + let message = "Live preview was interrupted. Retry to continue; recording is still running." + state = .failed(message) + statusMessage = message + } + } + + /// Runs what one track recorded since the last look through speech + /// detection and the turn cutter. Returns whether there was anything. + /// The track's state moves only when the whole look succeeded, so a + /// failed one is heard again from the same frame. + private func hear(track index: Int, folder: URL, session: Int) async throws -> Bool { + let chunkFrames = SpeechActivity.liveChunkFrames + let samples = try await audioPreparationWorker.readChunks( + source: folder.appendingPathComponent(tracks[index].fileName), + from: tracks[index].processedFrames, chunkFrames: chunkFrames, maximumChunks: Self.chunksPerLook) + guard isCurrent(session), !samples.isEmpty else { return false } + var cutter = tracks[index].cutter ?? LiveTurnCutter(startingAt: tracks[index].processedFrames) + var voice = tracks[index].voice + var tail = tracks[index].tail + let tailFrames = Int(LiveTranscriptChunker.cutSearchSeconds * Self.turnSampleRate) + var closed: [LiveTurnCutter.Turn] = [] + for offset in stride(from: 0, to: samples.count, by: chunkFrames) { + let chunk = Array(samples[offset..<(offset + chunkFrames)]) + let verdict = try await voiceActivity(chunk, voice) + guard isCurrent(session) else { return false } + voice = verdict.state + tail.append(contentsOf: chunk) + if tail.count > tailFrames { tail.removeFirst(tail.count - tailFrames) } + let turn = cutter.consume(probability: verdict.probability) { from, to in + let search = Array(tail.suffix(Int(min(to - from, Int64(tail.count))))) + let quiet = LiveTranscriptChunker.refinedCutOffset(tail: search, sampleRate: Self.turnSampleRate) + return to - Int64(search.count) + Int64(quiet) + } + if let turn { closed.append(turn) } + } + tracks[index].cutter = cutter + tracks[index].voice = voice + tracks[index].tail = tail + tracks[index].processedFrames += Int64(samples.count) + pendingTurns.append(contentsOf: closed.map { PendingTurn(track: index, turn: $0) }) + return true + } + + private func decode(_ pending: PendingTurn, folder: URL, session: Int) async throws -> [Line] { + let track = tracks[pending.track] + guard let span = try await audioPreparationWorker.prepareSpan( + source: folder.appendingPathComponent(track.fileName), + start: pending.turn.start, end: pending.turn.end) else { return [] } + do { + guard isCurrent(session) else { + await audioPreparationWorker.removePreparedChunk(at: span) + return [] + } + isWorking = true + defer { if generation == session { isWorking = false } } + if !hasTranscribedOnce { statusMessage = "Preparing the transcription model…" } + let started = ContinuousClock.now + let transcript = try await transcribe(span, settings()) + await audioPreparationWorker.removePreparedChunk(at: span) + guard isCurrent(session) else { return [] } + lastDecodeSeconds = Self.seconds(since: started) + hasTranscribedOnce = true + statusMessage = nil + let origin = Double(pending.turn.start) / Self.turnSampleRate + return transcript.segments.compactMap { segment in + let text = segment.displayText + return text.isEmpty ? nil : Line(time: origin + segment.start, speaker: track.speaker, text: text) + } + } catch { + await audioPreparationWorker.removePreparedChunk(at: span) + throw error + } + } + + private func publish(_ fresh: [Line], for pending: PendingTurn) { + let speaker = tracks[pending.track].speaker + // What was shown of this turn while it was spoken gives way to its text. + let start = Double(pending.turn.start) / Self.turnSampleRate + partials.removeAll { $0.speaker == speaker && $0.time <= start } + guard !fresh.isEmpty else { return } + lines = Array(((lines + fresh).sorted { $0.time < $1.time }).suffix(Self.maxLines)) + } + + /// A turn that closed without text (too short to keep) leaves nothing to + /// replace what was shown of it. + private func dropFinishedPartials() { + guard !partials.isEmpty else { return } + for (index, track) in tracks.enumerated() + where track.cutter?.openStart == nil && !pendingTurns.contains(where: { $0.track == index }) { + partials.removeAll { $0.speaker == track.speaker } + } + } + + /// With nothing waiting for the model, and a model on this Mac that has + /// just answered quickly, shows what the longest-running open turn says + /// so far. Its failure costs nothing: the turn's own text follows when + /// the speaker pauses. A transcription server hears each turn once, when + /// it is finished: showing it sooner would upload the same audio again + /// every two seconds. + private func showTurnInProgress(folder: URL, session: Int) async throws -> Bool { + guard let last = lastDecodeSeconds, last <= timing.interimBudgetSeconds, + !settings().transcriptionModel.isRemote else { return false } + let every = Int64(timing.interimEverySeconds * Self.turnSampleRate) + let candidates = tracks.indices.compactMap { index -> (track: Int, start: Int64, end: Int64)? in + guard let cutter = tracks[index].cutter, let start = cutter.openStart, !cutter.isPausing, + cutter.processed - max(start, tracks[index].interimAt) >= every else { return nil } + return (index, start, cutter.processed) + } + guard let open = candidates.min(by: { $0.start < $1.start }) else { return false } + tracks[open.track].interimAt = open.end + let track = tracks[open.track] + guard let span = try await audioPreparationWorker.prepareSpan( + source: folder.appendingPathComponent(track.fileName), start: open.start, end: open.end) else { return false } + do { + guard isCurrent(session) else { + await audioPreparationWorker.removePreparedChunk(at: span) + return false + } + let started = ContinuousClock.now + let transcript = try await transcribe(span, settings()) + await audioPreparationWorker.removePreparedChunk(at: span) + guard isCurrent(session) else { return false } + lastDecodeSeconds = Self.seconds(since: started) + let text = transcript.segments.map(\.displayText).filter { !$0.isEmpty }.joined(separator: " ") + partials.removeAll { $0.speaker == track.speaker } + if !text.isEmpty { + let line = Line(time: Double(open.start) / Self.turnSampleRate, speaker: track.speaker, text: text) + partials = (partials + [line]).sorted { $0.time < $1.time } + } + return true + } catch is CancellationError { + await audioPreparationWorker.removePreparedChunk(at: span) + throw CancellationError() + } catch { + await audioPreparationWorker.removePreparedChunk(at: span) + lokalbotLog("live text so far failed track=\(track.fileName) error=\(error.localizedDescription)") + return false + } + } + + private static func seconds(since start: ContinuousClock.Instant) -> Double { + let elapsed = (ContinuousClock.now - start).components + return Double(elapsed.seconds) + Double(elapsed.attoseconds) / 1e18 + } + // MARK: - Fast live text /// Your name, so Muse spells it as you do. diff --git a/LokalBot/Services/LiveTurnCutter.swift b/LokalBot/Services/LiveTurnCutter.swift new file mode 100644 index 00000000..e5afa0cf --- /dev/null +++ b/LokalBot/Services/LiveTurnCutter.swift @@ -0,0 +1,106 @@ +import Foundation + +/// Cuts one live track into turns while it is being recorded. It hears only +/// the voice-activity model's verdict on each 256 ms chunk, and closes a turn +/// once the pause after it is long enough, so the turn can be transcribed +/// while the speaker is still quiet. No audio and no model: the caller runs +/// the model and reads the turn's samples. +/// +/// The rules are FluidAudio's streaming ones (`VadManager.processStreamingChunk`) +/// with two additions: a long turn closes at its first quiet chunk, and a +/// turn with no pause at all is handed over at a maximum length. +/// Benchmarks/LiveTurns measured the settings. +struct LiveTurnCutter: Equatable, Sendable { + struct Settings: Equatable, Sendable { + /// Frames the voice-activity model judges at a time. + var chunkFrames: Int64 = 4_096 + var sampleRate = 16_000.0 + /// A chunk at or above this probability is speech, as after a meeting. + var speechThreshold: Float = 0.85 + /// A chunk below this one is quiet. In between, nothing changes. + var quietThreshold: Float = 0.70 + /// Quiet chunks that close a turn: two are a pause of about half a second. + var quietChunksToClose = 2 + /// A turn this long closes at its first quiet chunk. Several voices + /// on one track leave shorter pauses between them than one voice + /// leaves inside a sentence, and a long turn has context to spare. + var eagerAfterSeconds: Double? = 4 + /// A turn with no pause is handed over at this length. + var maximumTurnSeconds = 15.0 + /// Audio kept before the first speech chunk and after the last. + var paddingSeconds = 0.1 + /// Less than one chunk of speech is what a length cut leaves behind + /// when the speaker stops right after it, not a turn. + var minimumTurnSeconds = 0.25 + } + + struct Turn: Equatable, Sendable { + /// Frames on the track's timeline. + var start: Int64 + var end: Int64 + /// Cut by length, not by a pause. + var forced: Bool + } + + var settings = Settings() + /// Frames heard so far. + private(set) var processed: Int64 = 0 + /// Where the turn being spoken starts, if someone is speaking. + private(set) var openStart: Int64? + /// Frames heard when the current pause began. + private var quietSince: Int64? + private var lastEnd: Int64 = 0 + + init(settings: Settings = Settings(), startingAt frame: Int64 = 0) { + self.settings = settings + processed = frame + lastEnd = frame + } + + /// The open turn has reached a pause that may close it. + var isPausing: Bool { quietSince != nil } + + /// Takes the verdict on the next chunk and returns the turn it closed. + /// `quietPoint` picks where to cut a turn that reached the maximum + /// length, between the two frames it is given; without it the cut is at + /// the end of the chunk. + mutating func consume(probability: Float, + quietPoint: (_ from: Int64, _ to: Int64) -> Int64 = { _, to in to }) -> Turn? { + processed += settings.chunkFrames + let padding = Int64(settings.paddingSeconds * settings.sampleRate) + if probability >= settings.speechThreshold { + quietSince = nil + if openStart == nil { + openStart = max(lastEnd, processed - settings.chunkFrames - padding) + } + } else if probability < settings.quietThreshold, let start = openStart { + let pauseStart = quietSince ?? processed + quietSince = pauseStart + let speechEnd = pauseStart - settings.chunkFrames + let quietChunks = Int((processed - pauseStart) / settings.chunkFrames) + 1 + if quietChunks >= chunksToClose(turnFrames: speechEnd - start) { + openStart = nil + quietSince = nil + return close(start: start, end: min(processed, speechEnd + padding), forced: false) + } + } + if let start = openStart, quietSince == nil, + Double(processed - start) / settings.sampleRate >= settings.maximumTurnSeconds { + let cut = min(processed, max(start, quietPoint(start, processed))) + openStart = cut + return close(start: start, end: cut, forced: true) + } + return nil + } + + private func chunksToClose(turnFrames: Int64) -> Int { + if let eager = settings.eagerAfterSeconds, Double(turnFrames) / settings.sampleRate >= eager { return 1 } + return max(1, settings.quietChunksToClose) + } + + private mutating func close(start: Int64, end: Int64, forced: Bool) -> Turn? { + guard Double(end - start) / settings.sampleRate >= settings.minimumTurnSeconds else { return nil } + lastEnd = end + return Turn(start: start, end: end, forced: forced) + } +} diff --git a/LokalBot/Services/MeetingRadar/MeetingRadar.swift b/LokalBot/Services/MeetingRadar/MeetingRadar.swift index 7d86c9f2..cd1054be 100644 --- a/LokalBot/Services/MeetingRadar/MeetingRadar.swift +++ b/LokalBot/Services/MeetingRadar/MeetingRadar.swift @@ -8,8 +8,8 @@ protocol RadarLineFeed: AnyObject { var linesPublisher: AnyPublisher<[LiveMeetingTranscriber.Line], Never> { get } var currentLines: [LiveMeetingTranscriber.Line] { get } var isRunning: Bool { get } - /// Lines arrive turn by turn, within a second of being said, rather than - /// in 12-second chunks, so passes can follow each turn. + /// Lines arrive turn by turn, about a second after being said, rather + /// than in windows of several seconds, so passes can follow each turn. var deliversTurns: Bool { get } func activatePreview() } @@ -17,7 +17,7 @@ protocol RadarLineFeed: AnyObject { extension LiveMeetingTranscriber: RadarLineFeed { var linesPublisher: AnyPublisher<[Line], Never> { $lines.eraseToAnyPublisher() } var currentLines: [Line] { lines } - var deliversTurns: Bool { engine == .muse } + var deliversTurns: Bool { engine != .local } func activatePreview() { activate() } } @@ -27,8 +27,9 @@ extension LiveMeetingTranscriber: RadarLineFeed { /// local search found in earlier meetings) to the Think model chosen in /// Settings → Models, under that model's own local or remote setting. /// Passes run about every 20 s, or after each turn when fast live text -/// streams the transcript from Muse Voice Transcribe. Speaker separation -/// stays on this Mac, and audio does too unless fast live text is on. +/// delivers turns: cut at pauses on this Mac, or streamed from Muse Voice +/// Transcribe. Speaker separation stays on this Mac, and audio does too +/// unless fast live text streams it to Muse. /// /// Radar starts only when the person presses Start Radar during a recording, /// and only once Live Radar is turned on in Settings. It stops with the diff --git a/LokalBot/Views/SettingsView.swift b/LokalBot/Views/SettingsView.swift index 8f64705e..be6a5eec 100644 --- a/LokalBot/Views/SettingsView.swift +++ b/LokalBot/Views/SettingsView.swift @@ -771,6 +771,12 @@ struct SettingsView: View { SettingsLabel("Language and vocabulary", help: app.settings.transcriptionLanguage.displayName) } + Toggle(isOn: $app.settings.turnByTurnLiveTextEnabled) { + SettingsLabel("Turn-by-turn live text", + help: "While the live transcript preview runs, cuts the microphone and call audio at pauses and transcribes each turn with your Transcribe model, so its text appears about a second after the speaker stops. Live Radar then reads after each turn. Without it, the preview transcribes a few seconds at a time.") + } + .accessibilityIdentifier("settings.turnByTurnLiveText") + .settingTarget("settings.turnByTurnLiveTextEnabled", selected: app.focusedSettingID) modelRow(.assistant, title: "Think model", value: ModelSettingsPresentation.assistantName(app.settings), help: "Writes the notes. Also used by Ask and the overnight review.") } diff --git a/LokalBotTests/LiveTurnCutterTests.swift b/LokalBotTests/LiveTurnCutterTests.swift new file mode 100644 index 00000000..c01bca36 --- /dev/null +++ b/LokalBotTests/LiveTurnCutterTests.swift @@ -0,0 +1,119 @@ +import XCTest +@testable import LokalBot + +final class LiveTurnCutterTests: XCTestCase { + private let chunk: Int64 = 4_096 + private let padding: Int64 = 1_600 + private let speech: Float = 0.95 + private let quiet: Float = 0.05 + private let unsure: Float = 0.75 + + /// Feeds one verdict per chunk and returns each turn with the number of + /// chunks heard when it closed. + private func run(_ cutter: inout LiveTurnCutter, _ verdicts: [Float]) -> [(turn: LiveTurnCutter.Turn, chunks: Int)] { + var closed: [(LiveTurnCutter.Turn, Int)] = [] + for (index, verdict) in verdicts.enumerated() { + if let turn = cutter.consume(probability: verdict) { closed.append((turn, index + 1)) } + } + return closed + } + + private func short() -> LiveTurnCutter { + var settings = LiveTurnCutter.Settings() + settings.eagerAfterSeconds = nil + return LiveTurnCutter(settings: settings) + } + + func testTwoQuietChunksCloseATurn() { + var cutter = short() + let closed = run(&cutter, [quiet, quiet, speech, speech, speech, quiet, quiet, quiet]) + XCTAssertEqual(closed.count, 1) + // Speech fills chunks 3 to 5; the turn keeps 0.1 s on either side. + XCTAssertEqual(closed.first?.turn, .init(start: 2 * chunk - padding, end: 5 * chunk + padding, forced: false)) + XCTAssertEqual(closed.first?.chunks, 7, "closed at the second quiet chunk, half a second into the pause") + XCTAssertNil(cutter.openStart) + } + + func testOneQuietChunkIsNotAPause() { + var cutter = short() + let closed = run(&cutter, [speech, speech, quiet, speech, speech, quiet, quiet]) + XCTAssertEqual(closed.map(\.turn), [.init(start: 0, end: 5 * chunk + padding, forced: false)]) + } + + func testAnUnsureChunkNeitherStartsATurnNorRestartsThePause() { + var cutter = short() + XCTAssertTrue(run(&cutter, [unsure, unsure]).isEmpty) + XCTAssertNil(cutter.openStart, "only a confident chunk starts a turn") + let closed = run(&cutter, [speech, quiet, unsure, quiet]) + XCTAssertEqual(closed.map(\.turn), [.init(start: 2 * chunk - padding, end: 3 * chunk + padding, forced: false)]) + XCTAssertEqual(closed.first?.chunks, 4, "the pause kept running through the unsure chunk") + } + + func testALongTurnClosesAtItsFirstQuietChunk() { + var cutter = LiveTurnCutter() + XCTAssertEqual(cutter.settings.eagerAfterSeconds, 4) + // 16 chunks are 4.1 s of speech: one quiet chunk is enough. + let long = run(&cutter, Array(repeating: speech, count: 16) + [quiet]) + XCTAssertEqual(long.map(\.turn), [.init(start: 0, end: 16 * chunk + padding, forced: false)]) + + // 15 chunks are 3.8 s: still two. + var other = LiveTurnCutter() + XCTAssertTrue(run(&other, Array(repeating: speech, count: 15) + [quiet]).isEmpty) + XCTAssertTrue(other.isPausing) + XCTAssertEqual(run(&other, [quiet]).map(\.turn), [.init(start: 0, end: 15 * chunk + padding, forced: false)]) + } + + func testATurnWithoutAPauseIsHandedOverAtTheMaximumLength() { + var cutter = short() + // 15 s are 58.6 chunks, so the 59th reaches the maximum. + let closed = run(&cutter, Array(repeating: speech, count: 62)) + XCTAssertEqual(closed.map(\.turn), [.init(start: 0, end: 59 * chunk, forced: true)]) + XCTAssertEqual(closed.first?.chunks, 59) + XCTAssertEqual(cutter.openStart, 59 * chunk, "the speaker is still talking") + let rest = run(&cutter, [quiet, quiet]) + XCTAssertEqual(rest.map(\.turn), [.init(start: 59 * chunk, end: 62 * chunk + padding, forced: false)]) + } + + func testALengthCutLandsWhereTheCallerFindsQuiet() { + var cutter = short() + var turn: LiveTurnCutter.Turn? + var asked: (from: Int64, to: Int64)? + for _ in 0..<59 { + turn = cutter.consume(probability: speech) { from, to in + asked = (from, to) + return to - 5_000 + } ?? turn + } + XCTAssertEqual(asked?.from, 0) + XCTAssertEqual(asked?.to, 59 * chunk) + XCTAssertEqual(turn, .init(start: 0, end: 59 * chunk - 5_000, forced: true)) + XCTAssertEqual(cutter.openStart, 59 * chunk - 5_000, "the next turn starts at the cut") + } + + func testAPauseRightAfterALengthCutLeavesNoStub() { + var cutter = short() + let closed = run(&cutter, Array(repeating: speech, count: 59) + [quiet, quiet, quiet]) + XCTAssertEqual(closed.count, 1, "0.1 s of padding after the cut is not a turn") + XCTAssertNil(cutter.openStart) + } + + func testTurnsNeverOverlap() { + var cutter = LiveTurnCutter() + var verdicts: [Float] = [] + for burst in [17, 3, 20, 1, 6] { verdicts += Array(repeating: speech, count: burst) + [quiet, quiet] } + let turns = run(&cutter, verdicts).map(\.turn) + XCTAssertEqual(turns.count, 5) + for (earlier, later) in zip(turns, turns.dropFirst()) { + XCTAssertLessThanOrEqual(earlier.end, later.start) + } + } + + func testFramesAreOnTheTrackTimelineWhenThePreviewStartsLate() { + var settings = LiveTurnCutter.Settings() + settings.eagerAfterSeconds = nil + var cutter = LiveTurnCutter(settings: settings, startingAt: 160_000) + let closed = run(&cutter, [speech, speech, quiet, quiet]) + XCTAssertEqual(closed.map(\.turn), [.init(start: 160_000, end: 160_000 + 2 * chunk + padding, forced: false)]) + XCTAssertEqual(cutter.processed, 160_000 + 4 * chunk) + } +} diff --git a/LokalBotTests/LiveTurnTextTests.swift b/LokalBotTests/LiveTurnTextTests.swift new file mode 100644 index 00000000..30b014c1 --- /dev/null +++ b/LokalBotTests/LiveTurnTextTests.swift @@ -0,0 +1,340 @@ +import AVFoundation +import XCTest +@testable import LokalBot + +/// Turn-by-turn live text on this Mac: the preview cuts each track at pauses +/// and transcribes one turn at a time. Speech detection and the model are +/// stand-ins; a tone is speech and digital silence is a pause. +@MainActor +final class LiveTurnTextTests: XCTestCase { + + private var root: URL! + private var transcriber: LiveMeetingTranscriber! + private let rate = 16_000.0 + + override func setUpWithError() throws { + root = FileManager.default.temporaryDirectory + .appendingPathComponent("live-turn-text-tests-\(UUID().uuidString)", isDirectory: true) + try FileManager.default.createDirectory(at: root, withIntermediateDirectories: true) + } + + override func tearDownWithError() throws { + transcriber?.stop() + try? FileManager.default.removeItem(at: root) + } + + func testEachPauseDeliversItsTurn() async throws { + // 2 s of speech, a pause, 5 s of speech, a pause. + try write(track(seconds: 12, speech: [(1, 3), (5, 10)]), to: AudioPreviewTee.micFileName) + var heard: [Double] = [] + start { url, _ in + heard.append(try Self.seconds(of: url)) + return Self.transcript("turn \(heard.count)") + } + XCTAssertEqual(transcriber.engine, .turns) + XCTAssertTrue(transcriber.deliversTurns, "Live Radar follows each turn") + try await waitUntil { self.transcriber.lines.count == 2 } + + XCTAssertEqual(transcriber.lines.map(\.text), ["turn 1", "turn 2"]) + XCTAssertEqual(transcriber.lines.map(\.speaker), ["local", "local"]) + // A turn starts up to one 256 ms chunk and 0.1 s of padding before its first word. + XCTAssertEqual(transcriber.lines[0].time, 1, accuracy: 0.4) + XCTAssertEqual(transcriber.lines[1].time, 5, accuracy: 0.4) + // The model was handed each turn alone, not a window across both. + XCTAssertEqual(heard.count, 2) + XCTAssertEqual(heard[0], 2, accuracy: 0.8) + XCTAssertEqual(heard[1], 5, accuracy: 0.8) + XCTAssertTrue(transcriber.partials.isEmpty) + XCTAssertNil(transcriber.errorMessage) + } + + func testWindowsStayTheDefault() async throws { + try write(track(seconds: 12, speech: [(1, 3)]), to: AudioPreviewTee.micFileName) + start(settings: AppSettings(), detectSpeech: { _ in true }, transcribe: { _, _ in Self.transcript("window") }) + XCTAssertEqual(transcriber.engine, .local) + XCTAssertFalse(transcriber.deliversTurns) + try await waitUntil { !self.transcriber.lines.isEmpty } + } + + func testBothTracksShareTheModelInTheOrderTheirTurnsClosed() async throws { + try write(track(seconds: 9, speech: [(1, 2.5)]), to: AudioPreviewTee.micFileName) + try write(track(seconds: 9, speech: [(4, 5.5)]), to: AudioPreviewTee.systemFileName) + var running = 0 + var overlapped = false + start { _, _ in + running += 1 + overlapped = overlapped || running > 1 + try await Task.sleep(for: .milliseconds(5)) + running -= 1 + return Self.transcript("said") + } + try await waitUntil { self.transcriber.lines.count == 2 } + XCTAssertEqual(transcriber.lines.map(\.speaker), ["local", "them"]) + XCTAssertEqual(transcriber.lines[1].time, 4, accuracy: 0.4) + XCTAssertFalse(overlapped, "one turn at a time") + } + + func testAFailedTurnIsTriedAgainAndAppearsOnce() async throws { + try write(track(seconds: 6, speech: [(1, 3)]), to: AudioPreviewTee.micFileName) + var calls = 0 + start { _, _ in + calls += 1 + if calls == 1 { throw TestError.unavailable } + return Self.transcript("second try") + } + try await waitUntil { self.transcriber.lines.count == 1 } + XCTAssertEqual(calls, 2) + XCTAssertEqual(transcriber.lines.map(\.text), ["second try"]) + XCTAssertTrue(transcriber.isRunning) + } + + func testRetryAfterGivingUpTranscribesTheTurnThatWasWaiting() async throws { + try write(track(seconds: 6, speech: [(1, 3)]), to: AudioPreviewTee.micFileName) + var calls = 0 + var broken = true + start { _, _ in + calls += 1 + if broken { throw TestError.unavailable } + return Self.transcript("kept") + } + try await waitUntil { self.transcriber.errorMessage != nil } + XCTAssertEqual(calls, 3) + XCTAssertFalse(transcriber.isRunning) + XCTAssertTrue(transcriber.lines.isEmpty) + + broken = false + transcriber.activate() + try await waitUntil { self.transcriber.lines.count == 1 } + XCTAssertEqual(transcriber.lines.map(\.text), ["kept"]) + XCTAssertEqual(transcriber.lines[0].time, 1, accuracy: 0.4) + } + + func testAPauseOfThePreviewLosesNoTurn() async throws { + try write(track(seconds: 6, speech: [(1, 3)]), to: AudioPreviewTee.micFileName) + var calls = 0 + var release: CheckedContinuation? + start { _, _ in + calls += 1 + if calls == 1 { await withCheckedContinuation { release = $0 } } + return Self.transcript("answer \(calls)") + } + try await waitUntil { release != nil } + transcriber.pause() + release?.resume() + try await Task.sleep(for: .milliseconds(30)) + XCTAssertTrue(transcriber.lines.isEmpty, "an answer for a paused preview is not shown") + + transcriber.activate() + try await waitUntil { self.transcriber.lines.count == 1 } + XCTAssertEqual(transcriber.lines.map(\.text), ["answer 2"]) + } + + func testSpeechDetectionFailingStopsThePreviewBeforeTheModelHearsAnything() async throws { + try write(track(seconds: 6, speech: [(1, 3)]), to: AudioPreviewTee.micFileName) + var calls = 0 + start(voiceActivity: { _, _ in throw TestError.unavailable }, transcribe: { _, _ in + calls += 1 + return Self.transcript("never") + }) + try await waitUntil { self.transcriber.errorMessage != nil } + XCTAssertEqual(calls, 0) + XCTAssertFalse(transcriber.isRunning) + } + + func testATurnStillBeingSpokenIsShownThenReplacedByItsText() async throws { + // A short turn, then one that is still going when the file ends. + let speech: [(Double, Double)] = [(0.5, 1.5), (3, 11)] + try write(track(seconds: 11, speech: speech), to: AudioPreviewTee.micFileName) + var calls = 0 + start { _, _ in + calls += 1 + return Self.transcript("answer \(calls)") + } + try await waitUntil { self.transcriber.lines.count == 1 && self.transcriber.partials.count == 1 } + XCTAssertEqual(transcriber.lines.map(\.text), ["answer 1"]) + XCTAssertEqual(transcriber.partials.first?.speaker, "local") + XCTAssertEqual(transcriber.partials.first?.time ?? 0, 3, accuracy: 0.4) + XCTAssertEqual(transcriber.partials.first?.text, "answer 2") + + // The speaker stops: the turn's own text takes the partial's place. + try write(track(seconds: 13, speech: speech), to: AudioPreviewTee.micFileName) + try await waitUntil { self.transcriber.lines.count == 2 } + XCTAssertTrue(transcriber.partials.isEmpty) + XCTAssertEqual(transcriber.lines[1].time, 3, accuracy: 0.4) + } + + func testAModelThatAnswersSlowlyShowsOnlyFinishedTurns() async throws { + try write(track(seconds: 11, speech: [(0.5, 1.5), (3, 11)]), to: AudioPreviewTee.micFileName) + var calls = 0 + var timing = LiveMeetingTranscriber.Timing(idlePoll: .milliseconds(2), retryBaseSeconds: 0.002, + turnPoll: .milliseconds(2)) + timing.interimBudgetSeconds = 0.01 + start(timing: timing) { _, _ in + calls += 1 + try await Task.sleep(for: .milliseconds(40)) + return Self.transcript("answer \(calls)") + } + try await waitUntil { self.transcriber.lines.count == 1 } + try await Task.sleep(for: .milliseconds(150)) + XCTAssertTrue(transcriber.partials.isEmpty) + XCTAssertEqual(calls, 1, "the open turn was not sent to a model too slow to keep up") + } + + func testATranscriptionServerHearsEachTurnOnce() async throws { + try write(track(seconds: 11, speech: [(0.5, 1.5), (3, 11)]), to: AudioPreviewTee.micFileName) + var settings = Self.turnsOn + settings.transcriptionModel = .remoteServer + var calls = 0 + start(settings: settings) { _, _ in + calls += 1 + return Self.transcript("answer \(calls)") + } + try await waitUntil { self.transcriber.lines.count == 1 } + try await Task.sleep(for: .milliseconds(150)) + XCTAssertTrue(transcriber.partials.isEmpty) + XCTAssertEqual(calls, 1, "a turn still being spoken is not uploaded") + } + + // MARK: - Audio preparation worker + + func testWorkerHandsOverWholeChunksOnly() async throws { + let source = root.appendingPathComponent(AudioPreviewTee.micFileName) + let worker = LiveMeetingAudioPreparationWorker(storageRoot: root) + var samples = [Float](repeating: 0, count: 10_000) + samples[4_096] = 0.5 + try write(samples, to: AudioPreviewTee.micFileName) + + let missing = try await worker.readChunks(source: root.appendingPathComponent("absent.caf"), + from: 0, chunkFrames: 4_096, maximumChunks: 8) + XCTAssertTrue(missing.isEmpty) + let all = try await worker.readChunks(source: source, from: 0, chunkFrames: 4_096, maximumChunks: 8) + XCTAssertEqual(all.count, 8_192, "the last 1,808 frames wait for the rest of their chunk") + XCTAssertEqual(all[4_096], 0.5) + let one = try await worker.readChunks(source: source, from: 0, chunkFrames: 4_096, maximumChunks: 1) + XCTAssertEqual(one.count, 4_096) + let later = try await worker.readChunks(source: source, from: 4_096, chunkFrames: 4_096, maximumChunks: 8) + XCTAssertEqual(later.count, 4_096) + XCTAssertEqual(later[0], 0.5) + let none = try await worker.readChunks(source: source, from: 8_192, chunkFrames: 4_096, maximumChunks: 8) + XCTAssertTrue(none.isEmpty) + } + + func testWorkerWritesExactlyTheTurnAndRefusesAnotherSampleRate() async throws { + let source = root.appendingPathComponent(AudioPreviewTee.micFileName) + let worker = LiveMeetingAudioPreparationWorker(storageRoot: root) + try write((0..<32_000).map { Float($0) / 32_000 }, to: AudioPreviewTee.micFileName) + + let prepared = try await worker.prepareSpan(source: source, start: 8_000, end: 24_000) + let span = try XCTUnwrap(prepared) + let file = try AVAudioFile(forReading: span) + XCTAssertEqual(file.length, 16_000) + let buffer = try XCTUnwrap(AVAudioPCMBuffer(pcmFormat: file.processingFormat, frameCapacity: 1)) + try file.read(into: buffer, frameCount: 1) + XCTAssertEqual(try XCTUnwrap(buffer.floatChannelData)[0][0], 0.25, accuracy: 0.0001) + await worker.removePreparedChunk(at: span) + XCTAssertFalse(FileManager.default.fileExists(atPath: span.path)) + + let beyond = try await worker.prepareSpan(source: source, start: 40_000, end: 48_000) + XCTAssertNil(beyond) + + try write([Float](repeating: 0, count: 44_100), to: AudioPreviewTee.systemFileName, sampleRate: 44_100) + do { + _ = try await worker.readChunks(source: root.appendingPathComponent(AudioPreviewTee.systemFileName), + from: 0, chunkFrames: 4_096, maximumChunks: 1) + XCTFail("frames are counted at 16 kHz") + } catch {} + } + + // MARK: - Helpers + + private enum TestError: Error { case unavailable, timeout } + + private static let fast = LiveMeetingTranscriber.Timing(idlePoll: .milliseconds(2), retryBaseSeconds: 0.002, + turnPoll: .milliseconds(2)) + + /// A tone is speech; anything quieter is a pause. + private static let toneIsSpeech: LiveMeetingTranscriber.VoiceActivity = { chunk, state in + let energy = chunk.reduce(Float(0)) { $0 + $1 * $1 } / Float(max(1, chunk.count)) + return (energy.squareRoot() > 0.01 ? 0.95 : 0.02, state) + } + + private static var turnsOn: AppSettings { + var settings = AppSettings() + settings.turnByTurnLiveTextEnabled = true + return settings + } + + /// Starts a preview from the beginning of the recording: turn by turn, + /// polling every 2 ms, with a tone for speech, unless told otherwise. + private func start(settings: AppSettings? = nil, + timing: LiveMeetingTranscriber.Timing? = nil, + detectSpeech: LiveMeetingTranscriber.SpeechDetector? = nil, + voiceActivity: LiveMeetingTranscriber.VoiceActivity? = nil, + transcribe: @escaping LiveMeetingTranscriber.Transcribe) { + let settings = settings ?? Self.turnsOn + transcriber = LiveMeetingTranscriber(storageRoot: root, detectSpeech: detectSpeech, transcribe: transcribe, + voiceActivity: voiceActivity ?? Self.toneIsSpeech, + timing: timing ?? Self.fast) { settings } + transcriber.prepare(folder: root) + transcriber.activate(from: .beginning) + } + + private static func transcript(_ text: String) -> Transcript { + Transcript(segments: [.init(start: 0, end: 0.001, speaker: "speaker", text: text, confidence: nil)], + engine: "test") + } + + private static func seconds(of url: URL) throws -> Double { + let file = try AVAudioFile(forReading: url) + return Double(file.length) / file.processingFormat.sampleRate + } + + /// Silence with a 400 Hz tone wherever someone speaks. + private func track(seconds: Double, speech: [(Double, Double)]) -> [Float] { + var samples = [Float](repeating: 0, count: Int(seconds * rate)) + for (from, to) in speech { + for index in Int(from * rate).. Bool) async throws { + let deadline = ContinuousClock.now.advanced(by: .seconds(4)) + while !predicate() { + guard ContinuousClock.now < deadline else { + XCTFail("timed out waiting for live text: lines=\(transcriber.lines.map(\.text)) " + + "partials=\(transcriber.partials.map(\.text)) error=\(transcriber.errorMessage ?? "none")") + throw TestError.timeout + } + try await Task.sleep(for: .milliseconds(5)) + } + } +} diff --git a/PRIVACY.md b/PRIVACY.md index 8875bc65..3f36d539 100644 --- a/PRIVACY.md +++ b/PRIVACY.md @@ -390,7 +390,8 @@ The app may make these outbound connections: Meta's. Even then, nothing is read until you start it: - **Live Radar** runs only after you press Start Radar during a recording, for that recording, and turns on the live transcript preview if it is - off. About every 20 seconds (or after each turn with fast live text) it + off. About every 20 seconds (or after each turn, with turn-by-turn live + text or fast live text) it gives the Think model the preview lines that arrived since the previous request, up to 24 earlier lines of the same meeting, the meeting title, invited names from @@ -1387,7 +1388,12 @@ platform differences: the library. - **The live transcript preview** runs only when you start it during a recording. It uses the local transcription model and is not saved; the - full transcript replaces it after the meeting. + full transcript replaces it after the meeting. **Turn-by-turn live text** + (Settings → Meetings → Processing, off by default) makes it cut the audio + at pauses on this Mac and transcribe one turn at a time, so a turn's text + follows its speaker by about a second and Live Radar reads after each + turn. It changes when text appears, not where audio goes: the preview + still uses the Transcribe model you chose. - **Calendar** is off until you turn on **Use calendar to improve detection** in Settings → Meetings. Instead of the Mac Calendar it reads: - private calendar links you add (an iCal feed such as Google Calendar's From 725e265e10f95c0eb5f7a55e4e1924762f0d52c3 Mon Sep 17 00:00:00 2001 From: stevyhacker Date: Tue, 10 Nov 2026 22:24:51 +0100 Subject: [PATCH 2/2] Benchmarks: live transcript turn by turn, against windows and streaming LiveTurnLatencyBenchmarkTests replays the bundled sample call, in English and Serbian, through the app's engines as if it were being recorded: turn by turn with LiveTurnCutter, in the preview's windows, through FluidAudio's streaming recognizer, and in real time through LiveMeetingTranscriber itself. Benchmarks/LiveTurns has the scripts, the method and the results of 10 November 2026. On an M4 Max with one track per side, a turn's text followed its last word by a median 0.5 to 0.7 s (0.9 s at worst) with Parakeet, Granite fast and Qwen3-ASR 1.7B, and 1.3 to 1.4 s with Whisper, against 1.7 to 3.2 s (3.8 to 5.0 s at worst) for windows, with fewer word errors and no more CPU time. The three tests are opt-in and listed in unit-skips.json. --- Benchmarks/LiveTurns/README.md | 55 + Benchmarks/LiveTurns/fetch_stream_model.py | 47 + .../LiveTurns/results/2026-11-10/REPORT.md | 95 + .../results/2026-11-10/pause-rules.json | 2164 +++++++++++++++++ .../LiveTurns/results/2026-11-10/scores.json | 1831 ++++++++++++++ Benchmarks/LiveTurns/score.py | 298 +++ Benchmarks/LiveTurns/split_tracks.py | 50 + Benchmarks/README.md | 8 + .../LiveTurnLatencyBenchmarkTests.swift | 671 +++++ Scripts/ci/unit-skips.json | 8 + 10 files changed, 5227 insertions(+) create mode 100644 Benchmarks/LiveTurns/README.md create mode 100644 Benchmarks/LiveTurns/fetch_stream_model.py create mode 100644 Benchmarks/LiveTurns/results/2026-11-10/REPORT.md create mode 100644 Benchmarks/LiveTurns/results/2026-11-10/pause-rules.json create mode 100644 Benchmarks/LiveTurns/results/2026-11-10/scores.json create mode 100644 Benchmarks/LiveTurns/score.py create mode 100644 Benchmarks/LiveTurns/split_tracks.py create mode 100644 LokalBotTests/LiveTurnLatencyBenchmarkTests.swift diff --git a/Benchmarks/LiveTurns/README.md b/Benchmarks/LiveTurns/README.md new file mode 100644 index 00000000..2aedfe61 --- /dev/null +++ b/Benchmarks/LiveTurns/README.md @@ -0,0 +1,55 @@ +# Live transcript, turn by turn + +This harness measures how soon the live transcript preview can show what someone said, using only models on the Mac. It compares cutting each track at pauses and transcribing one turn at a time ("turns", the opt-in **Turn-by-turn live text** setting) with what the preview does by default, 4 to 12 second windows ("windows"), and with a streaming recognizer that the app links but does not use. It was built to find out whether fast live text needs Meta's Muse Voice Transcribe. + +Read the [results of 10 November 2026](results/2026-11-10/REPORT.md). + +## Protocol + +- **Recordings.** The bundled sample meeting, a scripted 3½-minute call between two synthetic voices, in English and Serbian (`LokalBot/Resources/SampleMeeting`). Its reference transcript has each turn's measured start and end. Each recording is used twice: as the one mixed track it ships as, which stands for a call track that carries several voices, and split into one track per speaker by `split_tracks.py`, the way the app records the microphone and the call audio. +- **Replay.** `LiveTurnLatencyBenchmarkTests` is an opt-in unit test, so it runs the app's own engines, speech detection and `LiveTurnCutter`. It hears each track one 256 ms chunk at a time, as live audio arrives. A turn's text is counted as ready at the recording time its pause was long enough, plus the time the model then took, with turns decoded one at a time across both tracks. The windows pass repeats `LiveMeetingTranscriber.run` on the same clock: a window is taken once 4 s are waiting, cut at a quiet point, and the preview looks again a second after it had nothing to do. Decoding is timed for real; only the recording's clock is simulated, and each recording gets one second of silence at its end so the last turn can close. +- **Hand-overs.** The delay is reported for every reference turn after which the other person speaks, 25 per recording: the time from the turn's last word to its text. A turn that follows the same speaker's previous one is not a hand-over. +- **Accuracy.** Word error rate of all pieces in order against the reference text, lower-cased with punctuation removed. Cyrillic output is transliterated before scoring, and its share is reported. Numbers are not normalized, so a model that writes "forty-two" for "42" is charged for it in both passes alike. +- **Streaming.** `testStreamingRecognizerDelay` feeds each turn to FluidAudio's `StreamingNemotronMultilingualAsrManager` 160 ms at a time and records when each token came out against the time it was spoken. +- **Real time.** `testPreviewReplayInRealTime` writes the two tracks to disk at the speed they were recorded and runs `LiveMeetingTranscriber` on them with the real speech detection and model, as in a meeting. It reports when each line reached the preview. + +## Reproduce + +Build the test bundle optimized; an unoptimized build decodes slower. Keep the build and the audio outside `~/Documents`, or macOS asks the freshly built test host for access to that folder and the run stalls. + +```sh +OUT=/private/tmp/lokalbot-live-turns +mkdir -p $OUT/audio +for L in en sr-Latn; do + cp LokalBot/Resources/SampleMeeting/orders-api.$L.m4a $OUT/audio/${L%-Latn}.m4a + python3 Benchmarks/LiveTurns/split_tracks.py $OUT/audio/${L%-Latn}.m4a \ + LokalBot/Resources/SampleMeeting/reference.$L.json $OUT/audio/${L%-Latn} +done +xcodegen generate +xcodebuild -project LokalBot.xcodeproj -scheme LokalBot -configuration Release -derivedDataPath $OUT/dd \ + -destination 'platform=macOS,arch=arm64' CODE_SIGNING_ALLOWED=NO ENABLE_TESTABILITY=YES ONLY_ACTIVE_ARCH=YES \ + 'OTHER_SWIFT_FLAGS=$(inherited) -DDEBUG -DLOKALBOT_TEST_HOOKS' -only-testing:LokalBotTests build-for-testing + +AUDIO="en=$OUT/audio/en.m4a|$OUT/audio/en.maya.wav|$OUT/audio/en.daniel.wav;sr=$OUT/audio/sr.m4a|$OUT/audio/sr.maya.wav|$OUT/audio/sr.daniel.wav" +run() { # run [VAR=value ...] + test=$1; shift + env TEST_RUNNER_LOKALBOT_LIVE_TURN_BENCH=$OUT/results TEST_RUNNER_LOKALBOT_LIVE_TURN_AUDIO="$AUDIO" "$@" \ + xcodebuild test-without-building -xctestrun $OUT/dd/Build/Products/*.xctestrun \ + -destination 'platform=macOS,arch=arm64' -only-testing:LokalBotTests/LiveTurnLatencyBenchmarkTests/$test +} +run testTurnByTurnDelay TEST_RUNNER_LOKALBOT_LIVE_TURN_MODELS=parakeet-v3,qwen3-asr-1.7b +run testTurnByTurnDelay TEST_RUNNER_LOKALBOT_LIVE_TURN_MODELS=parakeet-v3 TEST_RUNNER_LOKALBOT_LIVE_TURN_LANGUAGES=sr=hr +python3 Benchmarks/LiveTurns/fetch_stream_model.py multilingual/560ms $OUT/nemotron-multilingual-560 +run testStreamingRecognizerDelay TEST_RUNNER_LOKALBOT_LIVE_STREAM_MODEL=$OUT/nemotron-multilingual-560 \ + TEST_RUNNER_LOKALBOT_LIVE_TURN_LANGUAGES=en=en-US,sr=hr-HR +run testPreviewReplayInRealTime TEST_RUNNER_LOKALBOT_LIVE_TURN_MODELS=parakeet-v3 TEST_RUNNER_LOKALBOT_LIVE_REPLAY_SECONDS=90 +python3 Benchmarks/LiveTurns/score.py $OUT/results LokalBot/Resources/SampleMeeting --json scores.json +``` + +Models are named by their Settings identifiers (`TranscriptionModelChoice`), and only ones already downloaded run; nothing is fetched except by `fetch_stream_model.py`. `TEST_RUNNER_LOKALBOT_LIVE_TURN_RULES=2:4,2:off,1:off` times other pause rules (quiet chunks to close a turn, and the turn length from which one is enough) instead of the app's. A run with a language hint is saved beside the plain one, as `run--sr@hr.json`. Whisper takes about four minutes for both recordings; the others well under two. + +While the test host runs it is a second LokalBot in the Dock. Leave it alone: opening it makes the unsigned build ask for the Keychain. + +## Limits + +One scripted call per language, with synthetic voices, no noise, no overlapping speech and none of the other side's voice in the microphone track. Pauses between speakers are 0.5 to 0.85 s; people often leave less, and hesitate more inside a turn. The Serbian recording is voiced by multilingual voices, so its error rates say which model copes better, not how well any of them reads a native speaker. Everything ran on one Mac, an M4 Max; a slower Mac decodes slower, and the delay grows by that much. CPU time is the process's own, so work on the GPU and the Neural Engine is not in it. diff --git a/Benchmarks/LiveTurns/fetch_stream_model.py b/Benchmarks/LiveTurns/fetch_stream_model.py new file mode 100644 index 00000000..36193bab --- /dev/null +++ b/Benchmarks/LiveTurns/fetch_stream_model.py @@ -0,0 +1,47 @@ +#!/usr/bin/env python3 +"""Download one variant of FluidInference's Core ML conversion of NVIDIA Nemotron 3.5 ASR Streaming +(compiled models only, about 660 MB) into its own folder, for +LiveTurnLatencyBenchmarkTests/testStreamingRecognizerDelay. The app does not use this model. + +Usage: fetch_stream_model.py +""" +import json +import os +import sys +import urllib.request + +REPOSITORY = "FluidInference/Nemotron-3.5-ASR-Streaming-Multilingual-0.6b-CoreML" +REVISION = "a70b9d1c33594f1f4cf3f4eca9b4a0a8f44a9f59" +AGENT = {"User-Agent": "model-fetch"} + + +def main(): + variant, out = sys.argv[1], sys.argv[2] + url = f"https://huggingface.co/api/models/{REPOSITORY}/tree/{REVISION}/{variant}?recursive=1&limit=1000" + items = [] + while url: + with urllib.request.urlopen(urllib.request.Request(url, headers=AGENT), timeout=60) as response: + items += json.load(response) + link = response.headers.get("Link", "") + url = link.split(";")[0].strip("<> ") if 'rel="next"' in link else None + total = 0 + for item in items: + if item.get("type") != "file" or ".mlpackage" in item["path"]: + continue + target = os.path.join(out, item["path"][len(variant) + 1:]) + size = item.get("lfs", {}).get("size") or item.get("size", 0) + total += size + if os.path.exists(target) and os.path.getsize(target) == size: + continue + os.makedirs(os.path.dirname(target), exist_ok=True) + source = f"https://huggingface.co/{REPOSITORY}/resolve/{REVISION}/{item['path']}" + request = urllib.request.Request(source, headers=AGENT) + with urllib.request.urlopen(request, timeout=120) as response, open(target + ".part", "wb") as file: + while block := response.read(1 << 20): + file.write(block) + os.replace(target + ".part", target) + print(f"{variant}: {total / 1e6:.0f} MB in {out}") + + +if __name__ == "__main__": + main() diff --git a/Benchmarks/LiveTurns/results/2026-11-10/REPORT.md b/Benchmarks/LiveTurns/results/2026-11-10/REPORT.md new file mode 100644 index 00000000..77549829 --- /dev/null +++ b/Benchmarks/LiveTurns/results/2026-11-10/REPORT.md @@ -0,0 +1,95 @@ +# Live transcript, turn by turn, on this Mac — 10 November 2026 + +**Result: fast live text does not need Meta's API.** Cutting each track at pauses and transcribing one turn at a time shows a turn's text a median **0.5 to 0.7 s** after the speaker's last word with Parakeet, Granite fast and Qwen3-ASR 1.7B, and **0.9 s at worst**. Muse Voice Transcribe measured 0.74 s (median) on other audio. Whisper large-v3 turbo takes 1.3 to 1.4 s, 1.7 s at worst. The preview's windows, on the same recordings, take 1.7 to 3.2 s and 3.8 to 5.0 s at worst, and make more word errors because they cut mid-sentence. The app now has this as **Turn-by-turn live text** (Settings → Meetings → Processing, off by default). + +What it does not give is text while a long turn is still being spoken; the app covers that by transcribing the open turn again every two seconds when the model is fast enough. A streaming recognizer the app already links gives real word-by-word text in English (words 0.34 s behind the speaker, 4.7% word error) but reads Serbian badly (43.9% with a Croatian hint), so it is not proposed as the route for both languages. + +[How it was measured](../../README.md). One Apple M4 Max, 48 GB, macOS 27.2, the app's test bundle built optimized with Xcode 27.2. Models as the app pins them: Parakeet TDT 0.6B v3 and Parakeet Ultra 0.6B (Core ML, FluidAudio 0.17.7), Qwen3-ASR 1.7B MLX 8-bit (`e5450a2`), Whisper large-v3 turbo (WhisperKit 1.1.1, `v20240930`), Granite Speech 5 fast (`18ca3c1`), Silero VAD (Core ML). Granite Speech 4.1, Parakeet v2 and Qwen3-ASR 0.6B were not installed complete on this Mac and did not run. + +## A call as it is recorded: one track per side + +The sample call split into a microphone track and a call track, 25 hand-overs per recording. "Turns" closes a turn after two quiet 256 ms chunks, or one once the turn is 4 s long, and decodes turns one at a time. Times are from the last word of a turn to its text. + +| Model | Recording | Decoding a turn: median / longest | Turns: median / 9 in 10 / worst | Windows: median / 9 in 10 / worst | Word errors: turns / windows | +| --- | --- | --- | --- | --- | --- | +| Parakeet TDT 0.6B v3 | English | 0.04 / 0.06 s | **0.50** / 0.77 / 0.81 s | 1.86 / 3.36 / 3.83 s | 3.4% / 8.7% | +| Parakeet Ultra 0.6B | English | 0.04 / 0.06 s | **0.50** / 0.77 / 0.81 s | 2.23 / 3.66 / 4.26 s | 4.7% / 12.5% | +| Granite Speech 5 fast | English | 0.02 / 0.04 s | **0.48** / 0.75 / 0.79 s | 1.74 / 3.63 / 3.97 s | 11.2% / 13.9% | +| Qwen3-ASR 1.7B (the default) | English | 0.15 / 0.29 s | **0.64** / 0.83 / 0.86 s | 1.89 / 3.37 / 4.29 s | 2.9% / 8.5% | +| Whisper large-v3 turbo | English | 0.81 / 1.00 s | **1.32** / 1.49 / 1.52 s | 3.22 / 4.29 / 5.01 s | 5.8% / 7.6% | +| Qwen3-ASR 1.7B | Serbian | 0.17 / 0.57 s | **0.71** / 0.86 / 0.94 s | 2.23 / 3.73 / 4.15 s | 22.3% / 31.8% | +| Whisper large-v3 turbo | Serbian | 0.82 / 1.30 s | **1.37** / 1.58 / 1.71 s | 2.44 / 3.68 / 4.57 s | 15.6% / 24.2% | +| Parakeet Ultra 0.6B, Croatian hint | Serbian | 0.04 / 0.08 s | **0.57** / 0.70 / 0.72 s | 2.57 / 4.13 / 4.34 s | 25.8% / 32.4% | +| Parakeet TDT 0.6B v3, Croatian hint | Serbian | 0.05 / 0.08 s | **0.58** / 0.70 / 0.73 s | 2.07 / 3.48 / 4.30 s | 29.9% / 45.5% | + +- **Where the time goes.** Speech detection judges 256 ms at a time, so two quiet chunks are known 0.5 to 0.8 s after the last word, depending on where in a chunk the speaker stopped. That is nearly all of the delay for the fast models; Whisper adds about 0.8 s of decoding, which hardly changes with the length of the turn. +- **Windows.** A window waits until 4 s of a track are recorded, so a turn's last words sit in an unfinished window for up to 4 s. The last few seconds of a recording never reach a window while it is being recorded: in most runs the last hand-over got no text at all, and it is left out of the window columns. Window edges depend on how long each decode took, so their error rates move between runs (Parakeet Ultra's English windows scored 7.2% in an earlier run and 12.5% here). +- **Cost.** Turns took no more processor time than windows: 0.4 against 0.6 CPU seconds per minute of audio for Parakeet, 1.1 to 1.7 against 1.3 to 1.9 for Qwen, 1.2 to 1.5 against 1.4 to 2.1 for Whisper. Speech detection takes 0.15 ms per chunk. GPU and Neural Engine work is not in these figures. +- **What turns lose.** A turn shows nothing until its pause, so a word waits longer on average: 3.3 to 4.6 s from being spoken to being shown, against 2.6 to 3.6 s in windows. This is why the app also shows the turn in progress (below). + +Muse Voice Transcribe, for comparison, delivered a turn's text 0.74 s (median), 0.83 s (95th percentile) and 1.75 s (worst) after the turn ended, over 2,014 turns of AMI meetings and a home network connection ([report](../../../MuseVoiceTranscribe/results/2026-10-27/REPORT.md)). That is other audio, so the comparison is of size, not a ranking. + +## Several voices on one track + +The same recordings as one mixed track, which is what the call side looks like with two remote people. Pauses between speakers are 0.5 to 0.85 s here, about the length of the pause the cutter waits for. + +| Rule for closing a turn | English: hand-overs caught, wait after the last word (median / worst) | Serbian | Pieces, and how many under 2 s (English / Serbian) | +| --- | --- | --- | --- | +| 1 quiet chunk | 25 of 25, 0.41 / 0.52 s | 25 of 25, 0.41 / 0.54 s | 61, 22 short / 51, 12 short | +| 2 quiet chunks | 18 of 25, 0.71 / 9.22 s | 15 of 25, 0.67 / 11.60 s | 25, 2 short / 29, 3 short | +| **2, or 1 once the turn is 4 s long (the app's)** | **23 of 25, 0.46 / 5.09 s** | **24 of 25, 0.50 / 5.97 s** | 38, 2 short / 41, 5 short | +| 2, or 1 once the turn is 3 s long | 23 of 25, 0.45 / 5.09 s | 25 of 25, 0.42 / 0.67 s | 38, 2 short / 43, 5 short | +| 3 quiet chunks | 6 of 25, 6.58 / 13.80 s | 5 of 25, 4.46 / 13.21 s | 15 / 18, none short | + +- Two quiet chunks alone miss a third of the hand-overs on a shared track: the next speaker starts before the pause is long enough, and the text waits for a later pause or the 15-second limit. One quiet chunk catches all of them but cuts the call into twice as many pieces, a third of them under 2 s. +- Closing a turn at its first quiet chunk once it is 4 s long catches 23 and 24 of 25 with few short pieces, and word error rates stay where they were (Qwen English 3.3% against 3.4% and 2.5%; Serbian 22.3% against 24.6% and 22.7%; Parakeet v3 English 2.9% against 2.5% and 3.1%). With one track per side it catches all 25, as every rule does. +- With a model, the shared track gave text a median 0.51 s (Parakeet v3), 0.64 s (Qwen) and 1.32 s (Whisper) after an English hand-over, and 5 to 7 s at worst for the one or two that were missed. + +## Through the app's own preview, in real time + +The first 90 s of the English call written to disk at the speed it was recorded, with `LiveMeetingTranscriber` running turn by turn as it does in a meeting: 11 hand-overs, 17 lines. + +| Model | Text after a hand-over's last word: median / worst | First text after a turn begins: median / worst | Times the turn in progress was shown | +| --- | --- | --- | --- | +| Parakeet TDT 0.6B v3 | 0.62 / 0.94 s | 2.15 / 7.07 s | 31 | +| Qwen3-ASR 1.7B | 0.70 / 0.96 s | 2.17 / 7.20 s | 31 | + +This adds what the simulation leaves out: the preview looks for new audio every 100 ms, and a finished turn can wait behind the turn in progress being shown. The 7 s is the call's first turn: nothing is shown early until the model has answered once and proved quick. + +## A streaming recognizer, for word-by-word text + +FluidAudio 0.17.7, which the app already links, has `StreamingNemotronMultilingualAsrManager` for a Core ML conversion of NVIDIA Nemotron 3.5 ASR Streaming 0.6B. The app neither uses nor downloads it. Measured with the 560 ms variant (`multilingual/560ms` at `a70b9d1`, 664 MB), each turn as its own stream: + +| Recording | A word is shown after it is spoken: median / 9 in 10 / worst | Text complete after a hand-over's last word: median / worst | Word errors | CPU seconds per minute of audio | +| --- | --- | --- | --- | --- | +| English (`en-US`) | 0.34 / 0.51 / 0.90 s | 0.48 / 0.79 s | 4.7% | 2.1 | +| Serbian with the Croatian hint (`hr-HR`) | 0.34 / 0.50 / 0.91 s | 0.54 / 0.69 s | 43.9% | 2.1 | + +It lists Croatian among its 40 locales and not Serbian. On this recording it is about twice as wrong as Whisper, Qwen or Parakeet, so it would serve English meetings only. + +## What each route would take + +| Route | Languages | Text while a turn is spoken | What it takes | +| --- | --- | --- | --- | +| **Turns with the Transcribe model** (built, opt-in) | Whatever the chosen model reads | The open turn again every 2 s, when the model answers within 0.25 s and runs on this Mac | Nothing to download. Mac only so far: Windows and Linux need the cutter and a chunk-at-a-time speech detector in `lokalbot-speech`. | +| **Nemotron 3.5 streaming through FluidAudio** | 40 locales; good in English, poor in Serbian here | Yes, 0.34 s behind | A 664 MB download with its own model entry and pinned files, a streaming path in the preview, and the model running for the whole meeting (0.8 GB). The Core ML repository's licence tag is `other`; NVIDIA's model card names OpenMDW-1.1. Neither was reviewed. | +| **sherpa-onnx streaming recognizer** | The same Nemotron 3.5 model as ONNX (475 MB, int8), an English Nemotron (464 MB), English Zipformer (70 MB, upper case, no punctuation) | Yes | The runtime is already shipped: the Mac's `libsherpa-onnx-c-api.dylib` exports the online recognizer functions and the port's `sherpa-onnx` 1.13.8 crate has `OnlineRecognizer`. Neither is called today. The route for word-by-word text on Windows and Linux. Sizes are from the project's release listing; nothing was downloaded or measured here. | +| **Apple SpeechAnalyzer** | On this Mac, `SpeechTranscriber` offers 45 locales with no Serbian, Croatian or Bosnian; `DictationTranscriber` offers 54, with Croatian and no Serbian | Yes | macOS 26 or later, and the app supports macOS 15. Mac only. Not measured here. | +| **Muse Voice Transcribe** (today's fast live text) | Meta's list | Yes | Sends meeting audio to api.meta.ai, about $0.36 per meeting hour. | + +## Serbian + +- Whisper read the Serbian recording best (15.6% turn by turn), then Qwen (22.3%), then Parakeet with a Croatian hint (25.8% and 29.9%). Qwen wrote about half of it in Cyrillic under the app's `sr` hint; the errors are counted after transliteration. +- **Parakeet with the app's own Serbian setting is unusable**: 81 to 84% word errors, all in Cyrillic. FluidAudio treats `sr` as a Cyrillic-script language, and Parakeet's closest language, Croatian, is written in Latin. With `hr`, or with no hint, the same audio scores 24 to 30%. This is how the app behaves today after a meeting too, not something this change introduced, and it is not fixed here. +- All of this is one recording with multilingual synthetic voices. It ranks the models on that recording; it does not say how well they read a native speaker. + +## Not measured + +- A real meeting. One scripted call per language has no overlapping speech, no noise, none of the other side in the microphone track, and none of the hesitations that make people pause mid-sentence. +- Any Mac but an M4 Max. The delay grows by however much slower a Mac decodes a turn; the turn in progress stops being shown when a decode takes more than 0.25 s. +- The setting in the running app with a microphone. The real-time replay runs the preview's own code with real models, but it writes the two tracks itself. +- What showing the turn in progress costs in processor time and battery. +- GPU and Neural Engine use, and memory per model. +- The streaming recognizer's other variants (1,120 ms and longer chunks, which NVIDIA reports as more accurate), and the sherpa-onnx and Apple routes. + +[Scores](scores.json), [pause rules compared with Parakeet v3 and Qwen](pause-rules.json). diff --git a/Benchmarks/LiveTurns/results/2026-11-10/pause-rules.json b/Benchmarks/LiveTurns/results/2026-11-10/pause-rules.json new file mode 100644 index 00000000..0972fcfa --- /dev/null +++ b/Benchmarks/LiveTurns/results/2026-11-10/pause-rules.json @@ -0,0 +1,2164 @@ +{ + "cuts": [ + { + "recording": "en", + "layout": "mixed", + "rule": "1 quiet", + "pieces": 61, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.414, + "p90": 0.492, + "max": 0.516, + "n": 25 + }, + "medianSeconds": 2.504, + "longestSeconds": 10.44, + "under2s": 22 + }, + { + "recording": "en", + "layout": "mixed", + "rule": "2 quiet, 1 after 3 s", + "pieces": 38, + "forced": 0, + "handovers": 25, + "caught": 23, + "wait": { + "median": 0.454, + "p90": 0.768, + "max": 5.086, + "n": 25 + }, + "medianSeconds": 4.936, + "longestSeconds": 11.72, + "under2s": 2 + }, + { + "recording": "en", + "layout": "mixed", + "rule": "2 quiet, 1 after 4 s", + "pieces": 38, + "forced": 0, + "handovers": 25, + "caught": 23, + "wait": { + "median": 0.462, + "p90": 0.768, + "max": 5.086, + "n": 25 + }, + "medianSeconds": 5.064, + "longestSeconds": 11.72, + "under2s": 2 + }, + { + "recording": "en", + "layout": "mixed", + "rule": "2 quiet, 1 after 6 s", + "pieces": 31, + "forced": 0, + "handovers": 25, + "caught": 21, + "wait": { + "median": 0.58, + "p90": 5.086, + "max": 8.964, + "n": 25 + }, + "medianSeconds": 6.6, + "longestSeconds": 12.488, + "under2s": 1 + }, + { + "recording": "en", + "layout": "mixed", + "rule": "2 quiet", + "pieces": 25, + "forced": 3, + "handovers": 25, + "caught": 18, + "wait": { + "median": 0.71, + "p90": 6.622, + "max": 9.22, + "n": 25 + }, + "medianSeconds": 8.074, + "longestSeconds": 15.11, + "under2s": 2 + }, + { + "recording": "en", + "layout": "mixed", + "rule": "3 quiet", + "pieces": 15, + "forced": 9, + "handovers": 25, + "caught": 6, + "wait": { + "median": 6.58, + "p90": 13.644, + "max": 13.804, + "n": 25 + }, + "medianSeconds": 14.274, + "longestSeconds": 15.31, + "under2s": 0 + }, + { + "recording": "en", + "layout": "mixed", + "rule": "4 quiet", + "pieces": 15, + "forced": 11, + "handovers": 25, + "caught": 4, + "wait": { + "median": 6.068, + "p90": 13.774, + "max": 13.9, + "n": 25 + }, + "medianSeconds": 14.592, + "longestSeconds": 15.31, + "under2s": 0 + }, + { + "recording": "en", + "layout": "tracks", + "rule": "1 quiet", + "pieces": 61, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.414, + "p90": 0.492, + "max": 0.516, + "n": 25 + }, + "medianSeconds": 2.504, + "longestSeconds": 10.44, + "under2s": 22 + }, + { + "recording": "en", + "layout": "tracks", + "rule": "2 quiet, 1 after 3 s", + "pieces": 39, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.454, + "p90": 0.734, + "max": 0.772, + "n": 25 + }, + "medianSeconds": 4.552, + "longestSeconds": 11.72, + "under2s": 2 + }, + { + "recording": "en", + "layout": "tracks", + "rule": "2 quiet, 1 after 4 s", + "pieces": 39, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.462, + "p90": 0.734, + "max": 0.772, + "n": 25 + }, + "medianSeconds": 4.808, + "longestSeconds": 11.72, + "under2s": 2 + }, + { + "recording": "en", + "layout": "tracks", + "rule": "2 quiet, 1 after 6 s", + "pieces": 33, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.58, + "p90": 0.748, + "max": 0.772, + "n": 25 + }, + "medianSeconds": 5.064, + "longestSeconds": 12.488, + "under2s": 1 + }, + { + "recording": "en", + "layout": "tracks", + "rule": "2 quiet", + "pieces": 30, + "forced": 1, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.67, + "p90": 0.748, + "max": 0.772, + "n": 25 + }, + "medianSeconds": 5.192, + "longestSeconds": 15.048, + "under2s": 2 + }, + { + "recording": "en", + "layout": "tracks", + "rule": "3 quiet", + "pieces": 28, + "forced": 3, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.926, + "p90": 1.004, + "max": 1.028, + "n": 25 + }, + "medianSeconds": 7.624, + "longestSeconds": 14.254, + "under2s": 2 + }, + { + "recording": "en", + "layout": "tracks", + "rule": "4 quiet", + "pieces": 28, + "forced": 3, + "handovers": 25, + "caught": 25, + "wait": { + "median": 1.182, + "p90": 1.26, + "max": 1.284, + "n": 25 + }, + "medianSeconds": 7.624, + "longestSeconds": 14.254, + "under2s": 2 + }, + { + "recording": "sr", + "layout": "mixed", + "rule": "1 quiet", + "pieces": 51, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.406, + "p90": 0.528, + "max": 0.54, + "n": 25 + }, + "medianSeconds": 3.272, + "longestSeconds": 14.536, + "under2s": 12 + }, + { + "recording": "sr", + "layout": "mixed", + "rule": "2 quiet, 1 after 3 s", + "pieces": 43, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.42, + "p90": 0.588, + "max": 0.666, + "n": 25 + }, + "medianSeconds": 4.296, + "longestSeconds": 14.536, + "under2s": 5 + }, + { + "recording": "sr", + "layout": "mixed", + "rule": "2 quiet, 1 after 4 s", + "pieces": 41, + "forced": 0, + "handovers": 25, + "caught": 24, + "wait": { + "median": 0.5, + "p90": 0.666, + "max": 5.97, + "n": 25 + }, + "medianSeconds": 4.296, + "longestSeconds": 14.536, + "under2s": 5 + }, + { + "recording": "sr", + "layout": "mixed", + "rule": "2 quiet, 1 after 6 s", + "pieces": 35, + "forced": 1, + "handovers": 25, + "caught": 21, + "wait": { + "median": 0.57, + "p90": 1.758, + "max": 5.97, + "n": 25 + }, + "medianSeconds": 5.064, + "longestSeconds": 13.904, + "under2s": 3 + }, + { + "recording": "sr", + "layout": "mixed", + "rule": "2 quiet", + "pieces": 29, + "forced": 5, + "handovers": 25, + "caught": 15, + "wait": { + "median": 0.666, + "p90": 7.184, + "max": 11.602, + "n": 25 + }, + "medianSeconds": 5.832, + "longestSeconds": 15.104, + "under2s": 3 + }, + { + "recording": "sr", + "layout": "mixed", + "rule": "3 quiet", + "pieces": 18, + "forced": 10, + "handovers": 25, + "caught": 5, + "wait": { + "median": 4.464, + "p90": 11.684, + "max": 13.21, + "n": 25 + }, + "medianSeconds": 14.173, + "longestSeconds": 15.054, + "under2s": 0 + }, + { + "recording": "sr", + "layout": "mixed", + "rule": "4 quiet", + "pieces": 16, + "forced": 14, + "handovers": 25, + "caught": 1, + "wait": { + "median": 5.928, + "p90": 13.194, + "max": 14.864, + "n": 25 + }, + "medianSeconds": 14.557, + "longestSeconds": 15.054, + "under2s": 0 + }, + { + "recording": "sr", + "layout": "tracks", + "rule": "1 quiet", + "pieces": 52, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.406, + "p90": 0.528, + "max": 0.54, + "n": 25 + }, + "medianSeconds": 3.144, + "longestSeconds": 14.536, + "under2s": 13 + }, + { + "recording": "sr", + "layout": "tracks", + "rule": "2 quiet, 1 after 3 s", + "pieces": 45, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.47, + "p90": 0.612, + "max": 0.666, + "n": 25 + }, + "medianSeconds": 4.296, + "longestSeconds": 14.536, + "under2s": 7 + }, + { + "recording": "sr", + "layout": "tracks", + "rule": "2 quiet, 1 after 4 s", + "pieces": 44, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.528, + "p90": 0.662, + "max": 0.676, + "n": 25 + }, + "medianSeconds": 4.296, + "longestSeconds": 14.536, + "under2s": 7 + }, + { + "recording": "sr", + "layout": "tracks", + "rule": "2 quiet, 1 after 6 s", + "pieces": 39, + "forced": 1, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.582, + "p90": 0.746, + "max": 0.796, + "n": 25 + }, + "medianSeconds": 4.296, + "longestSeconds": 13.904, + "under2s": 7 + }, + { + "recording": "sr", + "layout": "tracks", + "rule": "2 quiet", + "pieces": 38, + "forced": 2, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.662, + "p90": 0.784, + "max": 0.796, + "n": 25 + }, + "medianSeconds": 4.04, + "longestSeconds": 15.054, + "under2s": 8 + }, + { + "recording": "sr", + "layout": "tracks", + "rule": "3 quiet", + "pieces": 31, + "forced": 2, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.918, + "p90": 1.04, + "max": 1.052, + "n": 25 + }, + "medianSeconds": 5.576, + "longestSeconds": 15.054, + "under2s": 3 + }, + { + "recording": "sr", + "layout": "tracks", + "rule": "4 quiet", + "pieces": 28, + "forced": 2, + "handovers": 25, + "caught": 25, + "wait": { + "median": 1.174, + "p90": 1.296, + "max": 1.308, + "n": 25 + }, + "medianSeconds": 8.392, + "longestSeconds": 15.054, + "under2s": 2 + } + ], + "passes": [ + { + "model": "parakeet-v3", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.106, + "footprintMB": 1174.971, + "pieces": 38, + "forced": 0, + "decode": { + "median": 0.045, + "p90": 0.055, + "max": 0.06, + "n": 38 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.207, + "p90": 0.451, + "max": 0.455, + "n": 38 + }, + "delay": { + "median": 0.506, + "p90": 0.804, + "max": 5.146, + "n": 25 + }, + "missing": 0, + "age": 3.364, + "wer": 2.899, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.37 + }, + { + "model": "parakeet-v3", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.106, + "footprintMB": 1174.971, + "pieces": 25, + "forced": 3, + "decode": { + "median": 0.054, + "p90": 0.071, + "max": 0.118, + "n": 25 + }, + "decodePerAudioSecond": 0.007, + "afterPause": { + "median": 0.464, + "p90": 0.478, + "max": 0.723, + "n": 22 + }, + "delay": { + "median": 0.756, + "p90": 6.74, + "max": 9.284, + "n": 25 + }, + "missing": 0, + "age": 5.821, + "wer": 2.536, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.391 + }, + { + "model": "parakeet-v3", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "1 quiet", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.106, + "footprintMB": 1174.971, + "pieces": 61, + "forced": 0, + "decode": { + "median": 0.039, + "p90": 0.048, + "max": 0.064, + "n": 61 + }, + "decodePerAudioSecond": 0.012, + "afterPause": { + "median": 0.195, + "p90": 0.204, + "max": 0.22, + "n": 61 + }, + "delay": { + "median": 0.456, + "p90": 0.536, + "max": 0.555, + "n": 25 + }, + "missing": 0, + "age": 2.59, + "wer": 3.08, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.461 + }, + { + "model": "parakeet-v3", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 3 s", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.106, + "footprintMB": 1174.971, + "pieces": 38, + "forced": 0, + "decode": { + "median": 0.045, + "p90": 0.054, + "max": 0.061, + "n": 38 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.206, + "p90": 0.45, + "max": 0.453, + "n": 38 + }, + "delay": { + "median": 0.501, + "p90": 0.805, + "max": 5.138, + "n": 25 + }, + "missing": 0, + "age": 3.395, + "wer": 3.08, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.395 + }, + { + "model": "parakeet-v3", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 6 s", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.106, + "footprintMB": 1174.971, + "pieces": 31, + "forced": 0, + "decode": { + "median": 0.047, + "p90": 0.06, + "max": 0.063, + "n": 31 + }, + "decodePerAudioSecond": 0.007, + "afterPause": { + "median": 0.217, + "p90": 0.456, + "max": 0.458, + "n": 31 + }, + "delay": { + "median": 0.62, + "p90": 5.14, + "max": 9.023, + "n": 25 + }, + "missing": 0, + "age": 4.313, + "wer": 3.804, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.358 + }, + { + "model": "parakeet-v3", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.106, + "footprintMB": 1174.971, + "pieces": 39, + "forced": 0, + "decode": { + "median": 0.043, + "p90": 0.051, + "max": 0.058, + "n": 39 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.206, + "p90": 0.451, + "max": 0.453, + "n": 39 + }, + "delay": { + "median": 0.504, + "p90": 0.771, + "max": 0.81, + "n": 25 + }, + "missing": 0, + "age": 3.344, + "wer": 3.442, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.359 + }, + { + "model": "parakeet-v3", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.106, + "footprintMB": 1174.971, + "pieces": 30, + "forced": 1, + "decode": { + "median": 0.048, + "p90": 0.068, + "max": 0.112, + "n": 30 + }, + "decodePerAudioSecond": 0.008, + "afterPause": { + "median": 0.459, + "p90": 0.478, + "max": 0.524, + "n": 29 + }, + "delay": { + "median": 0.716, + "p90": 0.805, + "max": 0.83, + "n": 25 + }, + "missing": 0, + "age": 4.967, + "wer": 2.717, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.386 + }, + { + "model": "parakeet-v3", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "1 quiet", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.106, + "footprintMB": 1174.971, + "pieces": 61, + "forced": 0, + "decode": { + "median": 0.039, + "p90": 0.048, + "max": 0.06, + "n": 61 + }, + "decodePerAudioSecond": 0.012, + "afterPause": { + "median": 0.195, + "p90": 0.204, + "max": 0.216, + "n": 61 + }, + "delay": { + "median": 0.455, + "p90": 0.535, + "max": 0.556, + "n": 25 + }, + "missing": 0, + "age": 2.593, + "wer": 3.623, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.454 + }, + { + "model": "parakeet-v3", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 3 s", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.106, + "footprintMB": 1174.971, + "pieces": 39, + "forced": 0, + "decode": { + "median": 0.043, + "p90": 0.052, + "max": 0.06, + "n": 39 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.206, + "p90": 0.45, + "max": 0.452, + "n": 39 + }, + "delay": { + "median": 0.505, + "p90": 0.77, + "max": 0.811, + "n": 25 + }, + "missing": 0, + "age": 3.375, + "wer": 3.623, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.388 + }, + { + "model": "parakeet-v3", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 6 s", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.106, + "footprintMB": 1174.971, + "pieces": 33, + "forced": 0, + "decode": { + "median": 0.047, + "p90": 0.059, + "max": 0.065, + "n": 33 + }, + "decodePerAudioSecond": 0.008, + "afterPause": { + "median": 0.447, + "p90": 0.457, + "max": 0.459, + "n": 33 + }, + "delay": { + "median": 0.621, + "p90": 0.791, + "max": 0.811, + "n": 25 + }, + "missing": 0, + "age": 4.113, + "wer": 3.442, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.377 + }, + { + "model": "parakeet-v3", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.106, + "footprintMB": 1174.971, + "pieces": 72, + "forced": 0, + "decode": { + "median": 0.043, + "p90": 0.048, + "max": 0.082, + "n": 72 + }, + "decodePerAudioSecond": 0.012, + "afterPause": null, + "delay": { + "median": 1.807, + "p90": 3.423, + "max": 3.777, + "n": 25 + }, + "missing": 0, + "age": 2.661, + "wer": 12.5, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.614 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.047, + "footprintMB": 844.955, + "pieces": 41, + "forced": 0, + "decode": { + "median": 0.043, + "p90": 0.057, + "max": 0.246, + "n": 41 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.213, + "p90": 0.45, + "max": 0.457, + "n": 41 + }, + "delay": { + "median": 0.557, + "p90": 0.721, + "max": 6.216, + "n": 25 + }, + "missing": 0, + "age": 3.919, + "wer": 83.811, + "cyrillic": 0.995, + "cpuPerAudioMinute": 0.386 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.047, + "footprintMB": 844.955, + "pieces": 29, + "forced": 5, + "decode": { + "median": 0.045, + "p90": 0.062, + "max": 0.095, + "n": 29 + }, + "decodePerAudioSecond": 0.007, + "afterPause": { + "median": 0.456, + "p90": 0.473, + "max": 0.477, + "n": 24 + }, + "delay": { + "median": 0.715, + "p90": 7.245, + "max": 11.651, + "n": 25 + }, + "missing": 0, + "age": 5.802, + "wer": 85.041, + "cyrillic": 0.992, + "cpuPerAudioMinute": 0.309 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "mixed", + "kind": "turns", + "rule": "1 quiet", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.047, + "footprintMB": 844.955, + "pieces": 51, + "forced": 0, + "decode": { + "median": 0.042, + "p90": 0.054, + "max": 0.075, + "n": 51 + }, + "decodePerAudioSecond": 0.01, + "afterPause": { + "median": 0.198, + "p90": 0.21, + "max": 0.231, + "n": 51 + }, + "delay": { + "median": 0.449, + "p90": 0.574, + "max": 0.584, + "n": 25 + }, + "missing": 0, + "age": 3.445, + "wer": 82.172, + "cyrillic": 0.996, + "cpuPerAudioMinute": 0.425 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 3 s", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.047, + "footprintMB": 844.955, + "pieces": 43, + "forced": 0, + "decode": { + "median": 0.043, + "p90": 0.054, + "max": 0.078, + "n": 43 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.208, + "p90": 0.45, + "max": 0.454, + "n": 43 + }, + "delay": { + "median": 0.481, + "p90": 0.626, + "max": 0.704, + "n": 25 + }, + "missing": 0, + "age": 3.754, + "wer": 83.197, + "cyrillic": 0.996, + "cpuPerAudioMinute": 0.402 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 6 s", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.047, + "footprintMB": 844.955, + "pieces": 35, + "forced": 1, + "decode": { + "median": 0.046, + "p90": 0.063, + "max": 0.251, + "n": 35 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.448, + "p90": 0.457, + "max": 0.459, + "n": 34 + }, + "delay": { + "median": 0.62, + "p90": 1.799, + "max": 6.221, + "n": 25 + }, + "missing": 0, + "age": 4.609, + "wer": 85.656, + "cyrillic": 0.986, + "cpuPerAudioMinute": 0.405 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.047, + "footprintMB": 844.955, + "pieces": 44, + "forced": 0, + "decode": { + "median": 0.044, + "p90": 0.054, + "max": 0.074, + "n": 44 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.22, + "p90": 0.452, + "max": 0.459, + "n": 44 + }, + "delay": { + "median": 0.572, + "p90": 0.701, + "max": 0.722, + "n": 25 + }, + "missing": 0, + "age": 3.75, + "wer": 81.352, + "cyrillic": 0.996, + "cpuPerAudioMinute": 0.398 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.047, + "footprintMB": 844.955, + "pieces": 38, + "forced": 2, + "decode": { + "median": 0.045, + "p90": 0.061, + "max": 0.246, + "n": 38 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.456, + "p90": 0.472, + "max": 0.48, + "n": 36 + }, + "delay": { + "median": 0.704, + "p90": 0.827, + "max": 0.844, + "n": 25 + }, + "missing": 0, + "age": 4.902, + "wer": 83.811, + "cyrillic": 0.993, + "cpuPerAudioMinute": 0.44 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "turns", + "rule": "1 quiet", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.047, + "footprintMB": 844.955, + "pieces": 52, + "forced": 0, + "decode": { + "median": 0.04, + "p90": 0.053, + "max": 0.074, + "n": 52 + }, + "decodePerAudioSecond": 0.01, + "afterPause": { + "median": 0.196, + "p90": 0.209, + "max": 0.23, + "n": 52 + }, + "delay": { + "median": 0.448, + "p90": 0.572, + "max": 0.585, + "n": 25 + }, + "missing": 0, + "age": 3.399, + "wer": 80.943, + "cyrillic": 0.996, + "cpuPerAudioMinute": 0.426 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 3 s", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.047, + "footprintMB": 844.955, + "pieces": 45, + "forced": 0, + "decode": { + "median": 0.043, + "p90": 0.054, + "max": 0.073, + "n": 45 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.207, + "p90": 0.45, + "max": 0.456, + "n": 45 + }, + "delay": { + "median": 0.511, + "p90": 0.656, + "max": 0.704, + "n": 25 + }, + "missing": 0, + "age": 3.688, + "wer": 81.557, + "cyrillic": 0.996, + "cpuPerAudioMinute": 0.403 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 6 s", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.047, + "footprintMB": 844.955, + "pieces": 39, + "forced": 1, + "decode": { + "median": 0.044, + "p90": 0.059, + "max": 0.071, + "n": 39 + }, + "decodePerAudioSecond": 0.008, + "afterPause": { + "median": 0.447, + "p90": 0.456, + "max": 0.459, + "n": 38 + }, + "delay": { + "median": 0.622, + "p90": 0.791, + "max": 0.841, + "n": 25 + }, + "missing": 0, + "age": 4.462, + "wer": 82.992, + "cyrillic": 0.993, + "cpuPerAudioMinute": 0.378 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 0.365, + "firstDecodeSeconds": 0.047, + "footprintMB": 844.955, + "pieces": 75, + "forced": 0, + "decode": { + "median": 0.043, + "p90": 0.049, + "max": 0.213, + "n": 75 + }, + "decodePerAudioSecond": 0.013, + "afterPause": null, + "delay": { + "median": 2.06, + "p90": 4.122, + "max": 4.143, + "n": 24 + }, + "missing": 1, + "age": 2.633, + "wer": 82.172, + "cyrillic": 0.995, + "cpuPerAudioMinute": 0.641 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.341, + "footprintMB": 9252.068, + "pieces": 38, + "forced": 0, + "decode": { + "median": 0.136, + "p90": 0.211, + "max": 0.274, + "n": 38 + }, + "decodePerAudioSecond": 0.027, + "afterPause": { + "median": 0.336, + "p90": 0.502, + "max": 0.521, + "n": 38 + }, + "delay": { + "median": 0.618, + "p90": 0.843, + "max": 5.297, + "n": 25 + }, + "missing": 0, + "age": 3.477, + "wer": 3.261, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.726 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.341, + "footprintMB": 9252.068, + "pieces": 25, + "forced": 3, + "decode": { + "median": 0.201, + "p90": 0.364, + "max": 0.422, + "n": 25 + }, + "decodePerAudioSecond": 0.026, + "afterPause": { + "median": 0.605, + "p90": 0.728, + "max": 0.852, + "n": 22 + }, + "delay": { + "median": 0.843, + "p90": 7.044, + "max": 9.482, + "n": 25 + }, + "missing": 0, + "age": 6.02, + "wer": 3.442, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.756 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "1 quiet", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.341, + "footprintMB": 9252.068, + "pieces": 61, + "forced": 0, + "decode": { + "median": 0.085, + "p90": 0.154, + "max": 0.25, + "n": 61 + }, + "decodePerAudioSecond": 0.03, + "afterPause": { + "median": 0.241, + "p90": 0.31, + "max": 0.406, + "n": 61 + }, + "delay": { + "median": 0.547, + "p90": 0.608, + "max": 0.685, + "n": 25 + }, + "missing": 0, + "age": 2.675, + "wer": 2.536, + "cyrillic": 0.0, + "cpuPerAudioMinute": 1.167 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 3 s", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.341, + "footprintMB": 9252.068, + "pieces": 38, + "forced": 0, + "decode": { + "median": 0.141, + "p90": 0.214, + "max": 0.284, + "n": 38 + }, + "decodePerAudioSecond": 0.027, + "afterPause": { + "median": 0.321, + "p90": 0.49, + "max": 0.518, + "n": 38 + }, + "delay": { + "median": 0.596, + "p90": 0.844, + "max": 5.302, + "n": 25 + }, + "missing": 0, + "age": 3.512, + "wer": 3.261, + "cyrillic": 0.0, + "cpuPerAudioMinute": 1.126 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 6 s", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.341, + "footprintMB": 9252.068, + "pieces": 31, + "forced": 0, + "decode": { + "median": 0.16, + "p90": 0.274, + "max": 0.29, + "n": 31 + }, + "decodePerAudioSecond": 0.026, + "afterPause": { + "median": 0.441, + "p90": 0.541, + "max": 0.563, + "n": 31 + }, + "delay": { + "median": 0.707, + "p90": 5.306, + "max": 9.227, + "n": 25 + }, + "missing": 0, + "age": 4.464, + "wer": 3.442, + "cyrillic": 0.0, + "cpuPerAudioMinute": 1.083 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.341, + "footprintMB": 9252.068, + "pieces": 39, + "forced": 0, + "decode": { + "median": 0.136, + "p90": 0.202, + "max": 0.264, + "n": 39 + }, + "decodePerAudioSecond": 0.027, + "afterPause": { + "median": 0.34, + "p90": 0.504, + "max": 0.524, + "n": 39 + }, + "delay": { + "median": 0.628, + "p90": 0.824, + "max": 0.864, + "n": 25 + }, + "missing": 0, + "age": 3.457, + "wer": 2.899, + "cyrillic": 0.0, + "cpuPerAudioMinute": 1.082 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.341, + "footprintMB": 9252.068, + "pieces": 30, + "forced": 1, + "decode": { + "median": 0.154, + "p90": 0.285, + "max": 0.375, + "n": 30 + }, + "decodePerAudioSecond": 0.026, + "afterPause": { + "median": 0.564, + "p90": 0.682, + "max": 0.787, + "n": 29 + }, + "delay": { + "median": 0.792, + "p90": 0.918, + "max": 1.093, + "n": 25 + }, + "missing": 0, + "age": 5.131, + "wer": 3.442, + "cyrillic": 0.0, + "cpuPerAudioMinute": 1.064 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "1 quiet", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.341, + "footprintMB": 9252.068, + "pieces": 61, + "forced": 0, + "decode": { + "median": 0.086, + "p90": 0.154, + "max": 0.25, + "n": 61 + }, + "decodePerAudioSecond": 0.03, + "afterPause": { + "median": 0.242, + "p90": 0.31, + "max": 0.406, + "n": 61 + }, + "delay": { + "median": 0.546, + "p90": 0.609, + "max": 0.683, + "n": 25 + }, + "missing": 0, + "age": 2.679, + "wer": 2.536, + "cyrillic": 0.0, + "cpuPerAudioMinute": 1.18 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 3 s", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.341, + "footprintMB": 9252.068, + "pieces": 39, + "forced": 0, + "decode": { + "median": 0.132, + "p90": 0.204, + "max": 0.269, + "n": 39 + }, + "decodePerAudioSecond": 0.027, + "afterPause": { + "median": 0.33, + "p90": 0.503, + "max": 0.519, + "n": 39 + }, + "delay": { + "median": 0.609, + "p90": 0.825, + "max": 0.864, + "n": 25 + }, + "missing": 0, + "age": 3.489, + "wer": 2.899, + "cyrillic": 0.0, + "cpuPerAudioMinute": 1.111 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 6 s", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.341, + "footprintMB": 9252.068, + "pieces": 33, + "forced": 0, + "decode": { + "median": 0.153, + "p90": 0.269, + "max": 0.292, + "n": 33 + }, + "decodePerAudioSecond": 0.026, + "afterPause": { + "median": 0.484, + "p90": 0.545, + "max": 0.569, + "n": 33 + }, + "delay": { + "median": 0.687, + "p90": 0.867, + "max": 0.895, + "n": 25 + }, + "missing": 0, + "age": 4.254, + "wer": 3.261, + "cyrillic": 0.0, + "cpuPerAudioMinute": 1.098 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.341, + "footprintMB": 9252.068, + "pieces": 73, + "forced": 0, + "decode": { + "median": 0.089, + "p90": 0.121, + "max": 0.133, + "n": 73 + }, + "decodePerAudioSecond": 0.024, + "afterPause": null, + "delay": { + "median": 2.182, + "p90": 3.366, + "max": 3.873, + "n": 24 + }, + "missing": 1, + "age": 2.702, + "wer": 11.232, + "cyrillic": 0.0, + "cpuPerAudioMinute": 1.265 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.266, + "footprintMB": 6857.193, + "pieces": 41, + "forced": 0, + "decode": { + "median": 0.18, + "p90": 0.382, + "max": 0.57, + "n": 41 + }, + "decodePerAudioSecond": 0.039, + "afterPause": { + "median": 0.46, + "p90": 0.558, + "max": 0.726, + "n": 41 + }, + "delay": { + "median": 0.71, + "p90": 0.902, + "max": 6.352, + "n": 25 + }, + "missing": 0, + "age": 4.132, + "wer": 22.336, + "cyrillic": 0.465, + "cpuPerAudioMinute": 1.643 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.266, + "footprintMB": 6857.193, + "pieces": 29, + "forced": 5, + "decode": { + "median": 0.201, + "p90": 0.492, + "max": 0.538, + "n": 29 + }, + "decodePerAudioSecond": 0.037, + "afterPause": { + "median": 0.589, + "p90": 0.847, + "max": 0.95, + "n": 24 + }, + "delay": { + "median": 0.935, + "p90": 7.59, + "max": 12.121, + "n": 25 + }, + "missing": 0, + "age": 6.114, + "wer": 24.59, + "cyrillic": 0.353, + "cpuPerAudioMinute": 1.541 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "mixed", + "kind": "turns", + "rule": "1 quiet", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.266, + "footprintMB": 6857.193, + "pieces": 51, + "forced": 0, + "decode": { + "median": 0.137, + "p90": 0.298, + "max": 0.567, + "n": 51 + }, + "decodePerAudioSecond": 0.04, + "afterPause": { + "median": 0.293, + "p90": 0.454, + "max": 0.723, + "n": 51 + }, + "delay": { + "median": 0.594, + "p90": 0.767, + "max": 0.861, + "n": 25 + }, + "missing": 0, + "age": 3.637, + "wer": 22.746, + "cyrillic": 0.45, + "cpuPerAudioMinute": 1.71 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 3 s", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.266, + "footprintMB": 6857.193, + "pieces": 43, + "forced": 0, + "decode": { + "median": 0.173, + "p90": 0.315, + "max": 0.59, + "n": 43 + }, + "decodePerAudioSecond": 0.04, + "afterPause": { + "median": 0.419, + "p90": 0.557, + "max": 0.746, + "n": 43 + }, + "delay": { + "median": 0.68, + "p90": 0.819, + "max": 0.922, + "n": 25 + }, + "missing": 0, + "age": 3.972, + "wer": 22.131, + "cyrillic": 0.453, + "cpuPerAudioMinute": 1.651 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 6 s", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.266, + "footprintMB": 6857.193, + "pieces": 35, + "forced": 1, + "decode": { + "median": 0.18, + "p90": 0.424, + "max": 0.494, + "n": 35 + }, + "decodePerAudioSecond": 0.037, + "afterPause": { + "median": 0.531, + "p90": 0.596, + "max": 0.649, + "n": 34 + }, + "delay": { + "median": 0.81, + "p90": 1.938, + "max": 6.352, + "n": 25 + }, + "missing": 0, + "age": 4.85, + "wer": 22.746, + "cyrillic": 0.329, + "cpuPerAudioMinute": 1.59 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.266, + "footprintMB": 6857.193, + "pieces": 44, + "forced": 0, + "decode": { + "median": 0.167, + "p90": 0.32, + "max": 0.59, + "n": 44 + }, + "decodePerAudioSecond": 0.04, + "afterPause": { + "median": 0.468, + "p90": 0.597, + "max": 0.746, + "n": 44 + }, + "delay": { + "median": 0.72, + "p90": 0.861, + "max": 0.948, + "n": 25 + }, + "missing": 0, + "age": 3.971, + "wer": 22.336, + "cyrillic": 0.589, + "cpuPerAudioMinute": 1.67 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.266, + "footprintMB": 6857.193, + "pieces": 38, + "forced": 2, + "decode": { + "median": 0.177, + "p90": 0.432, + "max": 0.543, + "n": 38 + }, + "decodePerAudioSecond": 0.039, + "afterPause": { + "median": 0.58, + "p90": 0.831, + "max": 0.955, + "n": 36 + }, + "delay": { + "median": 0.896, + "p90": 1.188, + "max": 1.229, + "n": 25 + }, + "missing": 0, + "age": 5.157, + "wer": 24.18, + "cyrillic": 0.477, + "cpuPerAudioMinute": 1.625 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "turns", + "rule": "1 quiet", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.266, + "footprintMB": 6857.193, + "pieces": 52, + "forced": 0, + "decode": { + "median": 0.136, + "p90": 0.285, + "max": 0.566, + "n": 52 + }, + "decodePerAudioSecond": 0.04, + "afterPause": { + "median": 0.292, + "p90": 0.441, + "max": 0.722, + "n": 52 + }, + "delay": { + "median": 0.586, + "p90": 0.786, + "max": 0.938, + "n": 25 + }, + "missing": 0, + "age": 3.594, + "wer": 24.385, + "cyrillic": 0.543, + "cpuPerAudioMinute": 1.119 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 3 s", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.266, + "footprintMB": 6857.193, + "pieces": 45, + "forced": 0, + "decode": { + "median": 0.156, + "p90": 0.312, + "max": 0.567, + "n": 45 + }, + "decodePerAudioSecond": 0.039, + "afterPause": { + "median": 0.448, + "p90": 0.557, + "max": 0.723, + "n": 45 + }, + "delay": { + "median": 0.687, + "p90": 0.807, + "max": 0.94, + "n": 25 + }, + "missing": 0, + "age": 3.901, + "wer": 22.541, + "cyrillic": 0.589, + "cpuPerAudioMinute": 1.099 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 6 s", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.266, + "footprintMB": 6857.193, + "pieces": 39, + "forced": 1, + "decode": { + "median": 0.175, + "p90": 0.416, + "max": 0.546, + "n": 39 + }, + "decodePerAudioSecond": 0.038, + "afterPause": { + "median": 0.529, + "p90": 0.618, + "max": 0.702, + "n": 38 + }, + "delay": { + "median": 0.802, + "p90": 0.969, + "max": 0.994, + "n": 25 + }, + "missing": 0, + "age": 4.708, + "wer": 23.566, + "cyrillic": 0.525, + "cpuPerAudioMinute": 1.557 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 1.371, + "firstDecodeSeconds": 0.266, + "footprintMB": 6857.193, + "pieces": 73, + "forced": 0, + "decode": { + "median": 0.126, + "p90": 0.186, + "max": 0.238, + "n": 73 + }, + "decodePerAudioSecond": 0.034, + "afterPause": null, + "delay": { + "median": 2.057, + "p90": 3.549, + "max": 3.996, + "n": 25 + }, + "missing": 0, + "age": 2.768, + "wer": 30.123, + "cyrillic": 0.548, + "cpuPerAudioMinute": 1.856 + } + ], + "streams": [], + "replays": [] +} diff --git a/Benchmarks/LiveTurns/results/2026-11-10/scores.json b/Benchmarks/LiveTurns/results/2026-11-10/scores.json new file mode 100644 index 00000000..e47fa9ba --- /dev/null +++ b/Benchmarks/LiveTurns/results/2026-11-10/scores.json @@ -0,0 +1,1831 @@ +{ + "cuts": [ + { + "recording": "en", + "layout": "mixed", + "rule": "1 quiet", + "pieces": 61, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.414, + "p90": 0.492, + "max": 0.516, + "n": 25 + }, + "medianSeconds": 2.504, + "longestSeconds": 10.44, + "under2s": 22 + }, + { + "recording": "en", + "layout": "mixed", + "rule": "2 quiet, 1 after 3 s", + "pieces": 38, + "forced": 0, + "handovers": 25, + "caught": 23, + "wait": { + "median": 0.454, + "p90": 0.768, + "max": 5.086, + "n": 25 + }, + "medianSeconds": 4.936, + "longestSeconds": 11.72, + "under2s": 2 + }, + { + "recording": "en", + "layout": "mixed", + "rule": "2 quiet, 1 after 4 s", + "pieces": 38, + "forced": 0, + "handovers": 25, + "caught": 23, + "wait": { + "median": 0.462, + "p90": 0.768, + "max": 5.086, + "n": 25 + }, + "medianSeconds": 5.064, + "longestSeconds": 11.72, + "under2s": 2 + }, + { + "recording": "en", + "layout": "mixed", + "rule": "2 quiet, 1 after 6 s", + "pieces": 31, + "forced": 0, + "handovers": 25, + "caught": 21, + "wait": { + "median": 0.58, + "p90": 5.086, + "max": 8.964, + "n": 25 + }, + "medianSeconds": 6.6, + "longestSeconds": 12.488, + "under2s": 1 + }, + { + "recording": "en", + "layout": "mixed", + "rule": "2 quiet", + "pieces": 25, + "forced": 3, + "handovers": 25, + "caught": 18, + "wait": { + "median": 0.71, + "p90": 6.622, + "max": 9.22, + "n": 25 + }, + "medianSeconds": 8.074, + "longestSeconds": 15.11, + "under2s": 2 + }, + { + "recording": "en", + "layout": "mixed", + "rule": "3 quiet", + "pieces": 15, + "forced": 9, + "handovers": 25, + "caught": 6, + "wait": { + "median": 6.58, + "p90": 13.644, + "max": 13.804, + "n": 25 + }, + "medianSeconds": 14.274, + "longestSeconds": 15.31, + "under2s": 0 + }, + { + "recording": "en", + "layout": "mixed", + "rule": "4 quiet", + "pieces": 15, + "forced": 11, + "handovers": 25, + "caught": 4, + "wait": { + "median": 6.068, + "p90": 13.774, + "max": 13.9, + "n": 25 + }, + "medianSeconds": 14.592, + "longestSeconds": 15.31, + "under2s": 0 + }, + { + "recording": "en", + "layout": "tracks", + "rule": "1 quiet", + "pieces": 61, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.414, + "p90": 0.492, + "max": 0.516, + "n": 25 + }, + "medianSeconds": 2.504, + "longestSeconds": 10.44, + "under2s": 22 + }, + { + "recording": "en", + "layout": "tracks", + "rule": "2 quiet, 1 after 3 s", + "pieces": 39, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.454, + "p90": 0.734, + "max": 0.772, + "n": 25 + }, + "medianSeconds": 4.552, + "longestSeconds": 11.72, + "under2s": 2 + }, + { + "recording": "en", + "layout": "tracks", + "rule": "2 quiet, 1 after 4 s", + "pieces": 39, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.462, + "p90": 0.734, + "max": 0.772, + "n": 25 + }, + "medianSeconds": 4.808, + "longestSeconds": 11.72, + "under2s": 2 + }, + { + "recording": "en", + "layout": "tracks", + "rule": "2 quiet, 1 after 6 s", + "pieces": 33, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.58, + "p90": 0.748, + "max": 0.772, + "n": 25 + }, + "medianSeconds": 5.064, + "longestSeconds": 12.488, + "under2s": 1 + }, + { + "recording": "en", + "layout": "tracks", + "rule": "2 quiet", + "pieces": 30, + "forced": 1, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.67, + "p90": 0.748, + "max": 0.772, + "n": 25 + }, + "medianSeconds": 5.192, + "longestSeconds": 15.048, + "under2s": 2 + }, + { + "recording": "en", + "layout": "tracks", + "rule": "3 quiet", + "pieces": 28, + "forced": 3, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.926, + "p90": 1.004, + "max": 1.028, + "n": 25 + }, + "medianSeconds": 7.624, + "longestSeconds": 14.254, + "under2s": 2 + }, + { + "recording": "en", + "layout": "tracks", + "rule": "4 quiet", + "pieces": 28, + "forced": 3, + "handovers": 25, + "caught": 25, + "wait": { + "median": 1.182, + "p90": 1.26, + "max": 1.284, + "n": 25 + }, + "medianSeconds": 7.624, + "longestSeconds": 14.254, + "under2s": 2 + }, + { + "recording": "sr", + "layout": "mixed", + "rule": "1 quiet", + "pieces": 51, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.406, + "p90": 0.528, + "max": 0.54, + "n": 25 + }, + "medianSeconds": 3.272, + "longestSeconds": 14.536, + "under2s": 12 + }, + { + "recording": "sr", + "layout": "mixed", + "rule": "2 quiet, 1 after 3 s", + "pieces": 43, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.42, + "p90": 0.588, + "max": 0.666, + "n": 25 + }, + "medianSeconds": 4.296, + "longestSeconds": 14.536, + "under2s": 5 + }, + { + "recording": "sr", + "layout": "mixed", + "rule": "2 quiet, 1 after 4 s", + "pieces": 41, + "forced": 0, + "handovers": 25, + "caught": 24, + "wait": { + "median": 0.5, + "p90": 0.666, + "max": 5.97, + "n": 25 + }, + "medianSeconds": 4.296, + "longestSeconds": 14.536, + "under2s": 5 + }, + { + "recording": "sr", + "layout": "mixed", + "rule": "2 quiet, 1 after 6 s", + "pieces": 35, + "forced": 1, + "handovers": 25, + "caught": 21, + "wait": { + "median": 0.57, + "p90": 1.758, + "max": 5.97, + "n": 25 + }, + "medianSeconds": 5.064, + "longestSeconds": 13.904, + "under2s": 3 + }, + { + "recording": "sr", + "layout": "mixed", + "rule": "2 quiet", + "pieces": 29, + "forced": 5, + "handovers": 25, + "caught": 15, + "wait": { + "median": 0.666, + "p90": 7.184, + "max": 11.602, + "n": 25 + }, + "medianSeconds": 5.832, + "longestSeconds": 15.104, + "under2s": 3 + }, + { + "recording": "sr", + "layout": "mixed", + "rule": "3 quiet", + "pieces": 18, + "forced": 10, + "handovers": 25, + "caught": 5, + "wait": { + "median": 4.464, + "p90": 11.684, + "max": 13.21, + "n": 25 + }, + "medianSeconds": 14.173, + "longestSeconds": 15.054, + "under2s": 0 + }, + { + "recording": "sr", + "layout": "mixed", + "rule": "4 quiet", + "pieces": 16, + "forced": 14, + "handovers": 25, + "caught": 1, + "wait": { + "median": 5.928, + "p90": 13.194, + "max": 14.864, + "n": 25 + }, + "medianSeconds": 14.557, + "longestSeconds": 15.054, + "under2s": 0 + }, + { + "recording": "sr", + "layout": "tracks", + "rule": "1 quiet", + "pieces": 52, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.406, + "p90": 0.528, + "max": 0.54, + "n": 25 + }, + "medianSeconds": 3.144, + "longestSeconds": 14.536, + "under2s": 13 + }, + { + "recording": "sr", + "layout": "tracks", + "rule": "2 quiet, 1 after 3 s", + "pieces": 45, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.47, + "p90": 0.612, + "max": 0.666, + "n": 25 + }, + "medianSeconds": 4.296, + "longestSeconds": 14.536, + "under2s": 7 + }, + { + "recording": "sr", + "layout": "tracks", + "rule": "2 quiet, 1 after 4 s", + "pieces": 44, + "forced": 0, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.528, + "p90": 0.662, + "max": 0.676, + "n": 25 + }, + "medianSeconds": 4.296, + "longestSeconds": 14.536, + "under2s": 7 + }, + { + "recording": "sr", + "layout": "tracks", + "rule": "2 quiet, 1 after 6 s", + "pieces": 39, + "forced": 1, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.582, + "p90": 0.746, + "max": 0.796, + "n": 25 + }, + "medianSeconds": 4.296, + "longestSeconds": 13.904, + "under2s": 7 + }, + { + "recording": "sr", + "layout": "tracks", + "rule": "2 quiet", + "pieces": 38, + "forced": 2, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.662, + "p90": 0.784, + "max": 0.796, + "n": 25 + }, + "medianSeconds": 4.04, + "longestSeconds": 15.054, + "under2s": 8 + }, + { + "recording": "sr", + "layout": "tracks", + "rule": "3 quiet", + "pieces": 31, + "forced": 2, + "handovers": 25, + "caught": 25, + "wait": { + "median": 0.918, + "p90": 1.04, + "max": 1.052, + "n": 25 + }, + "medianSeconds": 5.576, + "longestSeconds": 15.054, + "under2s": 3 + }, + { + "recording": "sr", + "layout": "tracks", + "rule": "4 quiet", + "pieces": 28, + "forced": 2, + "handovers": 25, + "caught": 25, + "wait": { + "median": 1.174, + "p90": 1.296, + "max": 1.308, + "n": 25 + }, + "medianSeconds": 8.392, + "longestSeconds": 15.054, + "under2s": 2 + } + ], + "passes": [ + { + "model": "granite-speech-5-turbo", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.175, + "firstDecodeSeconds": 0.04, + "footprintMB": 7878.913, + "pieces": 38, + "forced": 0, + "decode": { + "median": 0.022, + "p90": 0.028, + "max": 0.037, + "n": 38 + }, + "decodePerAudioSecond": 0.005, + "afterPause": { + "median": 0.183, + "p90": 0.432, + "max": 0.433, + "n": 38 + }, + "delay": { + "median": 0.484, + "p90": 0.787, + "max": 5.113, + "n": 25 + }, + "missing": 0, + "age": 3.342, + "wer": 11.413, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.12 + }, + { + "model": "granite-speech-5-turbo", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.175, + "firstDecodeSeconds": 0.04, + "footprintMB": 7878.913, + "pieces": 39, + "forced": 0, + "decode": { + "median": 0.022, + "p90": 0.028, + "max": 0.036, + "n": 39 + }, + "decodePerAudioSecond": 0.005, + "afterPause": { + "median": 0.183, + "p90": 0.432, + "max": 0.433, + "n": 39 + }, + "delay": { + "median": 0.484, + "p90": 0.753, + "max": 0.791, + "n": 25 + }, + "missing": 0, + "age": 3.323, + "wer": 11.232, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.119 + }, + { + "model": "granite-speech-5-turbo", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 0.175, + "firstDecodeSeconds": 0.04, + "footprintMB": 7878.913, + "pieces": 69, + "forced": 0, + "decode": { + "median": 0.02, + "p90": 0.022, + "max": 0.112, + "n": 69 + }, + "decodePerAudioSecond": 0.006, + "afterPause": null, + "delay": { + "median": 1.742, + "p90": 3.628, + "max": 3.974, + "n": 23 + }, + "missing": 2, + "age": 2.641, + "wer": 13.949, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.187 + }, + { + "model": "parakeet-ultra", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.461, + "firstDecodeSeconds": 0.111, + "footprintMB": 2100.939, + "pieces": 38, + "forced": 0, + "decode": { + "median": 0.047, + "p90": 0.055, + "max": 0.064, + "n": 38 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.209, + "p90": 0.451, + "max": 0.455, + "n": 38 + }, + "delay": { + "median": 0.509, + "p90": 0.805, + "max": 5.148, + "n": 25 + }, + "missing": 0, + "age": 3.366, + "wer": 4.891, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.433 + }, + { + "model": "parakeet-ultra", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.461, + "firstDecodeSeconds": 0.111, + "footprintMB": 2100.939, + "pieces": 39, + "forced": 0, + "decode": { + "median": 0.045, + "p90": 0.052, + "max": 0.059, + "n": 39 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.207, + "p90": 0.451, + "max": 0.453, + "n": 39 + }, + "delay": { + "median": 0.504, + "p90": 0.772, + "max": 0.812, + "n": 25 + }, + "missing": 0, + "age": 3.346, + "wer": 4.71, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.397 + }, + { + "model": "parakeet-ultra", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 0.461, + "firstDecodeSeconds": 0.111, + "footprintMB": 2100.939, + "pieces": 73, + "forced": 0, + "decode": { + "median": 0.041, + "p90": 0.045, + "max": 0.191, + "n": 73 + }, + "decodePerAudioSecond": 0.012, + "afterPause": null, + "delay": { + "median": 2.231, + "p90": 3.657, + "max": 4.261, + "n": 25 + }, + "missing": 0, + "age": 2.652, + "wer": 12.5, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.556 + }, + { + "model": "parakeet-ultra", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.461, + "firstDecodeSeconds": 0.058, + "footprintMB": 1495.486, + "pieces": 41, + "forced": 0, + "decode": { + "median": 0.047, + "p90": 0.068, + "max": 0.087, + "n": 41 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.224, + "p90": 0.453, + "max": 0.459, + "n": 41 + }, + "delay": { + "median": 0.576, + "p90": 0.703, + "max": 6.038, + "n": 25 + }, + "missing": 0, + "age": 3.918, + "wer": 84.221, + "cyrillic": 0.984, + "cpuPerAudioMinute": 0.464 + }, + { + "model": "parakeet-ultra", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.461, + "firstDecodeSeconds": 0.058, + "footprintMB": 1495.486, + "pieces": 44, + "forced": 0, + "decode": { + "median": 0.047, + "p90": 0.061, + "max": 0.086, + "n": 44 + }, + "decodePerAudioSecond": 0.01, + "afterPause": { + "median": 0.236, + "p90": 0.454, + "max": 0.464, + "n": 44 + }, + "delay": { + "median": 0.575, + "p90": 0.702, + "max": 0.724, + "n": 25 + }, + "missing": 0, + "age": 3.756, + "wer": 84.221, + "cyrillic": 1.0, + "cpuPerAudioMinute": 0.48 + }, + { + "model": "parakeet-ultra", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 0.461, + "firstDecodeSeconds": 0.058, + "footprintMB": 1495.486, + "pieces": 77, + "forced": 0, + "decode": { + "median": 0.044, + "p90": 0.048, + "max": 0.197, + "n": 77 + }, + "decodePerAudioSecond": 0.013, + "afterPause": null, + "delay": { + "median": 2.817, + "p90": 4.126, + "max": 4.282, + "n": 24 + }, + "missing": 1, + "age": 2.667, + "wer": 82.787, + "cyrillic": 0.997, + "cpuPerAudioMinute": 0.633 + }, + { + "model": "parakeet-ultra", + "recording": "sr", + "optimized": true, + "language": "hr", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.54, + "firstDecodeSeconds": 0.113, + "footprintMB": 2548.674, + "pieces": 41, + "forced": 0, + "decode": { + "median": 0.046, + "p90": 0.067, + "max": 0.082, + "n": 41 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.223, + "p90": 0.452, + "max": 0.46, + "n": 41 + }, + "delay": { + "median": 0.568, + "p90": 0.703, + "max": 6.035, + "n": 25 + }, + "missing": 0, + "age": 3.916, + "wer": 24.59, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.398 + }, + { + "model": "parakeet-ultra", + "recording": "sr", + "optimized": true, + "language": "hr", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.54, + "firstDecodeSeconds": 0.113, + "footprintMB": 2548.674, + "pieces": 44, + "forced": 0, + "decode": { + "median": 0.044, + "p90": 0.054, + "max": 0.079, + "n": 44 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.225, + "p90": 0.452, + "max": 0.459, + "n": 44 + }, + "delay": { + "median": 0.575, + "p90": 0.702, + "max": 0.721, + "n": 25 + }, + "missing": 0, + "age": 3.752, + "wer": 25.82, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.406 + }, + { + "model": "parakeet-ultra", + "recording": "sr", + "optimized": true, + "language": "hr", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 0.54, + "firstDecodeSeconds": 0.113, + "footprintMB": 2548.674, + "pieces": 77, + "forced": 0, + "decode": { + "median": 0.043, + "p90": 0.049, + "max": 0.201, + "n": 77 + }, + "decodePerAudioSecond": 0.013, + "afterPause": null, + "delay": { + "median": 2.575, + "p90": 4.126, + "max": 4.338, + "n": 24 + }, + "missing": 1, + "age": 2.646, + "wer": 32.377, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.625 + }, + { + "model": "parakeet-v3", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.385, + "firstDecodeSeconds": 0.111, + "footprintMB": 1450.127, + "pieces": 38, + "forced": 0, + "decode": { + "median": 0.046, + "p90": 0.056, + "max": 0.063, + "n": 38 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.211, + "p90": 0.452, + "max": 0.454, + "n": 38 + }, + "delay": { + "median": 0.506, + "p90": 0.805, + "max": 5.148, + "n": 25 + }, + "missing": 0, + "age": 3.366, + "wer": 2.899, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.398 + }, + { + "model": "parakeet-v3", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.385, + "firstDecodeSeconds": 0.111, + "footprintMB": 1450.127, + "pieces": 39, + "forced": 0, + "decode": { + "median": 0.044, + "p90": 0.052, + "max": 0.06, + "n": 39 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.207, + "p90": 0.451, + "max": 0.454, + "n": 39 + }, + "delay": { + "median": 0.505, + "p90": 0.771, + "max": 0.811, + "n": 25 + }, + "missing": 0, + "age": 3.345, + "wer": 3.442, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.401 + }, + { + "model": "parakeet-v3", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 0.385, + "firstDecodeSeconds": 0.111, + "footprintMB": 1450.127, + "pieces": 74, + "forced": 0, + "decode": { + "median": 0.042, + "p90": 0.048, + "max": 0.081, + "n": 74 + }, + "decodePerAudioSecond": 0.013, + "afterPause": null, + "delay": { + "median": 1.857, + "p90": 3.361, + "max": 3.827, + "n": 24 + }, + "missing": 1, + "age": 2.664, + "wer": 8.696, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.61 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.385, + "firstDecodeSeconds": 0.05, + "footprintMB": 1457.252, + "pieces": 41, + "forced": 0, + "decode": { + "median": 0.045, + "p90": 0.057, + "max": 0.264, + "n": 41 + }, + "decodePerAudioSecond": 0.01, + "afterPause": { + "median": 0.213, + "p90": 0.451, + "max": 0.457, + "n": 41 + }, + "delay": { + "median": 0.557, + "p90": 0.721, + "max": 6.234, + "n": 25 + }, + "missing": 0, + "age": 3.922, + "wer": 83.811, + "cyrillic": 0.995, + "cpuPerAudioMinute": 0.464 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.385, + "firstDecodeSeconds": 0.05, + "footprintMB": 1457.252, + "pieces": 44, + "forced": 0, + "decode": { + "median": 0.044, + "p90": 0.052, + "max": 0.07, + "n": 44 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.219, + "p90": 0.452, + "max": 0.459, + "n": 44 + }, + "delay": { + "median": 0.574, + "p90": 0.701, + "max": 0.721, + "n": 25 + }, + "missing": 0, + "age": 3.749, + "wer": 81.352, + "cyrillic": 0.996, + "cpuPerAudioMinute": 0.382 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 0.385, + "firstDecodeSeconds": 0.05, + "footprintMB": 1457.252, + "pieces": 75, + "forced": 0, + "decode": { + "median": 0.043, + "p90": 0.049, + "max": 0.201, + "n": 75 + }, + "decodePerAudioSecond": 0.013, + "afterPause": null, + "delay": { + "median": 2.093, + "p90": 3.684, + "max": 4.297, + "n": 24 + }, + "missing": 1, + "age": 2.648, + "wer": 81.967, + "cyrillic": 0.999, + "cpuPerAudioMinute": 0.647 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "hr", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.539, + "firstDecodeSeconds": 0.122, + "footprintMB": 1394.142, + "pieces": 41, + "forced": 0, + "decode": { + "median": 0.045, + "p90": 0.063, + "max": 0.125, + "n": 41 + }, + "decodePerAudioSecond": 0.009, + "afterPause": { + "median": 0.219, + "p90": 0.452, + "max": 0.46, + "n": 41 + }, + "delay": { + "median": 0.563, + "p90": 0.705, + "max": 6.095, + "n": 25 + }, + "missing": 0, + "age": 3.917, + "wer": 32.172, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.394 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "hr", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 0.539, + "firstDecodeSeconds": 0.122, + "footprintMB": 1394.142, + "pieces": 44, + "forced": 0, + "decode": { + "median": 0.047, + "p90": 0.056, + "max": 0.082, + "n": 44 + }, + "decodePerAudioSecond": 0.01, + "afterPause": { + "median": 0.228, + "p90": 0.452, + "max": 0.461, + "n": 44 + }, + "delay": { + "median": 0.576, + "p90": 0.702, + "max": 0.725, + "n": 25 + }, + "missing": 0, + "age": 3.753, + "wer": 29.918, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.445 + }, + { + "model": "parakeet-v3", + "recording": "sr", + "optimized": true, + "language": "hr", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 0.539, + "firstDecodeSeconds": 0.122, + "footprintMB": 1394.142, + "pieces": 74, + "forced": 0, + "decode": { + "median": 0.044, + "p90": 0.051, + "max": 0.211, + "n": 74 + }, + "decodePerAudioSecond": 0.013, + "afterPause": null, + "delay": { + "median": 2.072, + "p90": 3.481, + "max": 4.296, + "n": 24 + }, + "missing": 1, + "age": 2.637, + "wer": 45.492, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.657 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 1.552, + "firstDecodeSeconds": 0.567, + "footprintMB": 6991.224, + "pieces": 38, + "forced": 0, + "decode": { + "median": 0.139, + "p90": 0.224, + "max": 0.275, + "n": 38 + }, + "decodePerAudioSecond": 0.028, + "afterPause": { + "median": 0.337, + "p90": 0.502, + "max": 0.525, + "n": 38 + }, + "delay": { + "median": 0.635, + "p90": 0.845, + "max": 5.299, + "n": 25 + }, + "missing": 0, + "age": 3.484, + "wer": 3.261, + "cyrillic": 0.0, + "cpuPerAudioMinute": 0.928 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 1.552, + "firstDecodeSeconds": 0.567, + "footprintMB": 6991.224, + "pieces": 39, + "forced": 0, + "decode": { + "median": 0.147, + "p90": 0.205, + "max": 0.295, + "n": 39 + }, + "decodePerAudioSecond": 0.029, + "afterPause": { + "median": 0.352, + "p90": 0.507, + "max": 0.524, + "n": 39 + }, + "delay": { + "median": 0.641, + "p90": 0.826, + "max": 0.864, + "n": 25 + }, + "missing": 0, + "age": 3.466, + "wer": 2.899, + "cyrillic": 0.0, + "cpuPerAudioMinute": 1.115 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 1.552, + "firstDecodeSeconds": 0.567, + "footprintMB": 6991.224, + "pieces": 70, + "forced": 0, + "decode": { + "median": 0.093, + "p90": 0.123, + "max": 0.135, + "n": 70 + }, + "decodePerAudioSecond": 0.025, + "afterPause": null, + "delay": { + "median": 1.889, + "p90": 3.371, + "max": 4.29, + "n": 24 + }, + "missing": 1, + "age": 2.713, + "wer": 8.514, + "cyrillic": 0.0, + "cpuPerAudioMinute": 1.32 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 1.552, + "firstDecodeSeconds": 0.262, + "footprintMB": 6990.083, + "pieces": 41, + "forced": 0, + "decode": { + "median": 0.181, + "p90": 0.382, + "max": 0.616, + "n": 41 + }, + "decodePerAudioSecond": 0.04, + "afterPause": { + "median": 0.456, + "p90": 0.571, + "max": 0.772, + "n": 41 + }, + "delay": { + "median": 0.714, + "p90": 0.922, + "max": 6.352, + "n": 25 + }, + "missing": 0, + "age": 4.141, + "wer": 22.336, + "cyrillic": 0.465, + "cpuPerAudioMinute": 1.734 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 1.552, + "firstDecodeSeconds": 0.262, + "footprintMB": 6990.083, + "pieces": 44, + "forced": 0, + "decode": { + "median": 0.168, + "p90": 0.314, + "max": 0.569, + "n": 44 + }, + "decodePerAudioSecond": 0.04, + "afterPause": { + "median": 0.465, + "p90": 0.593, + "max": 0.725, + "n": 44 + }, + "delay": { + "median": 0.71, + "p90": 0.857, + "max": 0.945, + "n": 25 + }, + "missing": 0, + "age": 3.968, + "wer": 22.336, + "cyrillic": 0.589, + "cpuPerAudioMinute": 1.725 + }, + { + "model": "qwen3-asr-1.7b", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 1.552, + "firstDecodeSeconds": 0.262, + "footprintMB": 6990.083, + "pieces": 72, + "forced": 0, + "decode": { + "median": 0.132, + "p90": 0.191, + "max": 0.24, + "n": 72 + }, + "decodePerAudioSecond": 0.034, + "afterPause": null, + "delay": { + "median": 2.23, + "p90": 3.727, + "max": 4.147, + "n": 24 + }, + "missing": 1, + "age": 2.765, + "wer": 31.762, + "cyrillic": 0.504, + "cpuPerAudioMinute": 1.86 + }, + { + "model": "whisper-large-v3-turbo", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 1.614, + "firstDecodeSeconds": 0.944, + "footprintMB": 2234.8, + "pieces": 38, + "forced": 0, + "decode": { + "median": 0.81, + "p90": 0.931, + "max": 1.015, + "n": 38 + }, + "decodePerAudioSecond": 0.157, + "afterPause": { + "median": 1.027, + "p90": 1.162, + "max": 1.194, + "n": 38 + }, + "delay": { + "median": 1.32, + "p90": 1.499, + "max": 6.017, + "n": 25 + }, + "missing": 0, + "age": 4.171, + "wer": 5.978, + "cyrillic": 0.0, + "cpuPerAudioMinute": 1.109 + }, + { + "model": "whisper-large-v3-turbo", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 1.614, + "firstDecodeSeconds": 0.944, + "footprintMB": 2234.8, + "pieces": 39, + "forced": 0, + "decode": { + "median": 0.813, + "p90": 0.897, + "max": 1.003, + "n": 39 + }, + "decodePerAudioSecond": 0.16, + "afterPause": { + "median": 1.022, + "p90": 1.169, + "max": 1.191, + "n": 39 + }, + "delay": { + "median": 1.318, + "p90": 1.487, + "max": 1.523, + "n": 25 + }, + "missing": 0, + "age": 4.146, + "wer": 5.797, + "cyrillic": 0.0, + "cpuPerAudioMinute": 1.166 + }, + { + "model": "whisper-large-v3-turbo", + "recording": "en", + "optimized": true, + "language": "en", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 1.614, + "firstDecodeSeconds": 0.944, + "footprintMB": 2234.8, + "pieces": 72, + "forced": 0, + "decode": { + "median": 0.75, + "p90": 0.8, + "max": 0.847, + "n": 72 + }, + "decodePerAudioSecond": 0.206, + "afterPause": null, + "delay": { + "median": 3.215, + "p90": 4.29, + "max": 5.006, + "n": 24 + }, + "missing": 1, + "age": 3.449, + "wer": 7.609, + "cyrillic": 0.0, + "cpuPerAudioMinute": 1.438 + }, + { + "model": "whisper-large-v3-turbo", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "mixed", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 1.996, + "firstDecodeSeconds": 1.159, + "footprintMB": 2201.003, + "pieces": 41, + "forced": 0, + "decode": { + "median": 0.86, + "p90": 1.08, + "max": 1.324, + "n": 41 + }, + "decodePerAudioSecond": 0.164, + "afterPause": { + "median": 1.119, + "p90": 1.292, + "max": 1.48, + "n": 41 + }, + "delay": { + "median": 1.377, + "p90": 1.718, + "max": 7.05, + "n": 25 + }, + "missing": 0, + "age": 4.838, + "wer": 16.598, + "cyrillic": 0.02, + "cpuPerAudioMinute": 1.754 + }, + { + "model": "whisper-large-v3-turbo", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "turns", + "rule": "2 quiet, 1 after 4 s", + "loadSeconds": 1.996, + "firstDecodeSeconds": 1.159, + "footprintMB": 2201.003, + "pieces": 44, + "forced": 0, + "decode": { + "median": 0.823, + "p90": 1.056, + "max": 1.3, + "n": 44 + }, + "decodePerAudioSecond": 0.17, + "afterPause": { + "median": 1.112, + "p90": 1.321, + "max": 1.456, + "n": 44 + }, + "delay": { + "median": 1.367, + "p90": 1.585, + "max": 1.712, + "n": 25 + }, + "missing": 0, + "age": 4.648, + "wer": 15.574, + "cyrillic": 0.02, + "cpuPerAudioMinute": 1.521 + }, + { + "model": "whisper-large-v3-turbo", + "recording": "sr", + "optimized": true, + "language": "sr", + "layout": "tracks", + "kind": "windows", + "rule": "windows", + "loadSeconds": 1.996, + "firstDecodeSeconds": 1.159, + "footprintMB": 2201.003, + "pieces": 72, + "forced": 0, + "decode": { + "median": 0.803, + "p90": 0.903, + "max": 1.889, + "n": 72 + }, + "decodePerAudioSecond": 0.23, + "afterPause": null, + "delay": { + "median": 2.435, + "p90": 3.681, + "max": 4.566, + "n": 24 + }, + "missing": 1, + "age": 3.558, + "wer": 24.18, + "cyrillic": 0.03, + "cpuPerAudioMinute": 2.121 + } + ], + "streams": [ + { + "model": "nemotron-multilingual-560", + "recording": "en", + "language": "en-US", + "optimized": true, + "chunkMilliseconds": 560, + "loadSeconds": 0.274, + "footprintMB": 806.439, + "pieces": 39, + "age": { + "median": 0.339, + "p90": 0.507, + "max": 0.9, + "n": 1309 + }, + "step": { + "median": 0.0, + "p90": 0.017, + "max": 0.34, + "n": 1394 + }, + "flush": { + "median": 0.015, + "p90": 0.019, + "max": 0.022, + "n": 39 + }, + "delay": { + "median": 0.477, + "p90": 0.756, + "max": 0.788, + "n": 25 + }, + "wer": 4.71, + "cyrillic": 0.0, + "cpuPerAudioMinute": 2.086 + }, + { + "model": "nemotron-multilingual-560", + "recording": "sr", + "language": "hr-HR", + "optimized": true, + "chunkMilliseconds": 560, + "loadSeconds": 0.274, + "footprintMB": 811.939, + "pieces": 44, + "age": { + "median": 0.336, + "p90": 0.5, + "max": 0.907, + "n": 1519 + }, + "step": { + "median": 0.0, + "p90": 0.017, + "max": 0.347, + "n": 1495 + }, + "flush": { + "median": 0.016, + "p90": 0.018, + "max": 0.026, + "n": 44 + }, + "delay": { + "median": 0.542, + "p90": 0.679, + "max": 0.69, + "n": 25 + }, + "wer": 43.852, + "cyrillic": 0.0, + "cpuPerAudioMinute": 2.135 + } + ], + "replays": [ + { + "model": "parakeet-v3", + "recording": "en", + "language": "en", + "optimized": true, + "seconds": 90, + "lines": 17, + "partials": 31, + "delay": { + "median": 0.624, + "p90": 0.943, + "max": 0.943, + "n": 11 + }, + "missing": 0, + "firstWords": { + "median": 2.149, + "p90": 2.4, + "max": 7.065, + "n": 14 + } + }, + { + "model": "qwen3-asr-1.7b", + "recording": "en", + "language": "en", + "optimized": true, + "seconds": 90, + "lines": 17, + "partials": 31, + "delay": { + "median": 0.696, + "p90": 0.931, + "max": 0.961, + "n": 11 + }, + "missing": 0, + "firstWords": { + "median": 2.173, + "p90": 2.743, + "max": 7.196, + "n": 14 + } + } + ] +} diff --git a/Benchmarks/LiveTurns/score.py b/Benchmarks/LiveTurns/score.py new file mode 100644 index 00000000..f6621617 --- /dev/null +++ b/Benchmarks/LiveTurns/score.py @@ -0,0 +1,298 @@ +#!/usr/bin/env python3 +"""Score the live-turn measurement (LokalBotTests/LiveTurnLatencyBenchmarkTests). + +For every reference turn that ends with a change of speaker (a hand-over), how long after its +last word did the text exist? Reported for turn-by-turn decoding, for the preview's windows, for +the streaming recognizer and for the real-time replay, with word error rate against the +reference transcript where the whole recording was decoded. + +Usage: score.py [--json out.json] +""" +import glob +import json +import os +import re +import statistics +import sys +import unicodedata + +REFERENCE = {"en": "en", "sr": "sr-Latn", "de": "de", "zh": "zh"} +CHUNK = 4096 / 16000 + +CYRILLIC = { + "а": "a", "б": "b", "в": "v", "г": "g", "д": "d", "ђ": "đ", "е": "e", "ж": "ž", "з": "z", "и": "i", + "ј": "j", "к": "k", "л": "l", "љ": "lj", "м": "m", "н": "n", "њ": "nj", "о": "o", "п": "p", "р": "r", + "с": "s", "т": "t", "ћ": "ć", "у": "u", "ф": "f", "х": "h", "ц": "c", "ч": "č", "џ": "dž", "ш": "š", +} + + +def words(text): + text = unicodedata.normalize("NFC", text).lower() + text = "".join(CYRILLIC.get(ch, ch) for ch in text) + text = re.sub(r"[^\w\s]", " ", text, flags=re.UNICODE).replace("_", " ") + return text.split() + + +def cyrillic_share(text): + letters = [ch for ch in text if ch.isalpha()] + return sum(1 for ch in letters if "Ѐ" <= ch <= "ӿ") / len(letters) if letters else 0.0 + + +def wer(reference, hypothesis): + r, h = words(reference), words(hypothesis) + previous = list(range(len(h) + 1)) + for i in range(1, len(r) + 1): + current = [i] + [0] * len(h) + for j in range(1, len(h) + 1): + current[j] = min(previous[j] + 1, current[j - 1] + 1, previous[j - 1] + (r[i - 1] != h[j - 1])) + previous = current + return 100.0 * previous[-1] / max(1, len(r)) + + +def quantile(values, q): + values = sorted(values) + return values[min(len(values) - 1, max(0, round(q * (len(values) - 1))))] if values else float("nan") + + +def summary(values): + if not values: + return {"median": float("nan"), "p90": float("nan"), "max": float("nan"), "n": 0} + return {"median": statistics.median(values), "p90": quantile(values, 0.9), "max": max(values), "n": len(values)} + + +def load_reference(sample, label): + name = REFERENCE.get(label.split("-")[0]) + path = os.path.join(sample, f"reference.{name}.json") + if not name or not os.path.exists(path): + return None + turns = json.load(open(path))["turns"] + order = [] + for turn in turns: + if turn["speaker"] not in order: + order.append(turn["speaker"]) + for index, turn in enumerate(turns): + turn["track"] = order.index(turn["speaker"]) + following = turns[index + 1] if index + 1 < len(turns) else None + turn["handsOver"] = following is None or following["speaker"] != turn["speaker"] + turn["nextStart"] = following["start"] if following else float("inf") + return turns + + +def covering(pieces, turn, layout, slack): + """The piece that holds the turn's last word, on the turn's own track.""" + track = turn["track"] if layout == "tracks" else 0 + mine = [p for p in pieces if p["track"] == track] + for piece in mine: + if piece["start"] - slack <= turn["end"] <= piece["end"] + slack: + return piece + later = [p for p in mine if p["end"] >= turn["end"]] + return min(later, key=lambda p: p["end"]) if later and slack > 0 else None + + +def mean_age(pieces): + """Average age of the audio in a piece when its text appears, weighted by length.""" + total = sum(p["end"] - p["start"] for p in pieces) + if total <= 0: + return float("nan") + return sum((p["ready"] - (p["start"] + p["end"]) / 2) * (p["end"] - p["start"]) for p in pieces) / total + + +def rule_label(rule): + if not rule: + return "windows" + eager = rule.get("eagerAfterSeconds") + return f"{rule['quietChunks']} quiet" + (f", 1 after {eager:g} s" if eager else "") + + +def rounded(value): + """Milliseconds are enough, and a missing number is null, which every JSON reader takes.""" + if isinstance(value, float): + return None if value != value else round(value, 3) + if isinstance(value, dict): + return {key: rounded(item) for key, item in value.items()} + if isinstance(value, list): + return [rounded(item) for item in value] + return value + + +def main(): + results, sample = sys.argv[1], sys.argv[2] + out = {"cuts": [], "passes": []} + + for path in sorted(glob.glob(os.path.join(results, "cut-*.json"))): + cut = json.load(open(path)) + reference = load_reference(sample, cut["recording"]) + if not reference: + continue + waits, caught, handovers = [], 0, 0 + for turn in reference: + if not turn["handsOver"]: + continue + handovers += 1 + piece = covering(cut["pieces"], turn, cut["layout"], 0.35) + if piece is None: + continue + waits.append(piece["decision"] - turn["end"]) + if not piece["forced"] and piece["end"] <= turn["nextStart"] - 0.05: + caught += 1 + lengths = [p["end"] - p["start"] for p in cut["pieces"]] + out["cuts"].append({ + "recording": cut["recording"], "layout": cut["layout"], "rule": rule_label(cut["rule"]), + "pieces": len(cut["pieces"]), "forced": sum(1 for p in cut["pieces"] if p["forced"]), + "handovers": handovers, "caught": caught, "wait": summary(waits), + "medianSeconds": statistics.median(lengths), "longestSeconds": max(lengths), + "under2s": sum(1 for x in lengths if x < 2), + }) + + for path in sorted(glob.glob(os.path.join(results, "run-*.json"))): + run = json.load(open(path)) + reference = load_reference(sample, run["recording"]) + if not reference: + continue + reference_text = " ".join(t["text"] for t in reference) + for item in run["passes"]: + pieces = item["pieces"] + windows = item["kind"] == "windows" + delays, missing = [], 0 + for turn in reference: + if not turn["handsOver"]: + continue + piece = covering(pieces, turn, item["layout"], 0.0 if windows else 0.35) + if piece is None or piece["start"] > turn["end"]: + missing += 1 + else: + delays.append(piece["ready"] - turn["end"]) + text = " ".join(p["text"] for p in sorted(pieces, key=lambda p: (p["start"], p["track"]))) + audio = sum(p["end"] - p["start"] for p in pieces) + decode = [p["decodeSeconds"] for p in pieces] + closed = [p for p in pieces if not p["forced"]] + out["passes"].append({ + "model": run["model"], "recording": run["recording"], "optimized": run["optimized"], + "language": run["language"], "layout": item["layout"], "kind": item["kind"], "rule": rule_label(item.get("rule")), + "loadSeconds": run["loadSeconds"], "firstDecodeSeconds": run["firstDecodeSeconds"], + "footprintMB": run["footprintMegabytes"], "pieces": len(pieces), + "forced": sum(1 for p in pieces if p["forced"]), + "decode": summary(decode), "decodePerAudioSecond": sum(decode) / audio if audio else float("nan"), + "afterPause": summary([p["ready"] - p["end"] for p in closed]) if not windows else None, + "delay": summary(delays), "missing": missing, "age": mean_age(pieces), + "wer": wer(reference_text, text), "cyrillic": cyrillic_share(text), + "cpuPerAudioMinute": 60 * sum(p["cpuSeconds"] for p in pieces) / (run["audioSeconds"] or 1), + }) + + out["streams"] = [] + for path in sorted(glob.glob(os.path.join(results, "stream-*.json"))): + run = json.load(open(path)) + reference = load_reference(sample, run["recording"]) + if not reference: + continue + pieces = run["pieces"] + delays = [] + for turn in reference: + if turn["handsOver"]: + piece = covering(pieces, turn, run["layout"], 0.35) + if piece is not None: + delays.append(piece["ready"] - turn["end"]) + text = " ".join(p["text"] for p in sorted(pieces, key=lambda p: (p["start"], p["track"]))) + out["streams"].append({ + "model": run["model"], "recording": run["recording"], "language": run["language"], + "optimized": run["optimized"], "chunkMilliseconds": run["chunkMilliseconds"], + "loadSeconds": run["loadSeconds"], "footprintMB": run["footprintMegabytes"], + "pieces": len(pieces), "age": summary(run["tokenAges"]), "step": summary(run["stepSeconds"]), + "flush": summary([p["decodeSeconds"] for p in pieces]), "delay": summary(delays), + "wer": wer(" ".join(t["text"] for t in reference), text), "cyrillic": cyrillic_share(text), + "cpuPerAudioMinute": 60 * sum(p["cpuSeconds"] for p in pieces) / (run["audioSeconds"] or 1), + }) + + out["replays"] = [] + for path in sorted(glob.glob(os.path.join(results, "replay-*.json"))): + run = json.load(open(path)) + reference = load_reference(sample, run["recording"]) + if not reference: + continue + speakers = ["local", "them"] + lines = [a for a in run["arrivals"] if not a["partial"]] + delays, first, missing = [], [], 0 + for turn in reference: + if turn["end"] > run["seconds"] - 0.5: + continue + mine = [a for a in run["arrivals"] if a["speaker"] == speakers[turn["track"]]] + begun = [a["at"] for a in mine if turn["start"] - 0.5 <= a["time"] <= turn["end"]] + if begun: + first.append(min(begun) - turn["start"]) + if not turn["handsOver"]: + continue + said = [a for a in mine if not a["partial"] and a["time"] <= turn["end"]] + last = max(said, key=lambda a: a["time"]) if said else None + if last is None or last["at"] < turn["end"]: + missing += 1 + else: + delays.append(last["at"] - turn["end"]) + out["replays"].append({ + "model": run["model"], "recording": run["recording"], "language": run["language"], + "optimized": run["optimized"], "seconds": run["seconds"], "lines": len(lines), + "partials": len(run["arrivals"]) - len(lines), "delay": summary(delays), "missing": missing, + "firstWords": summary(first), + }) + + if "--json" in sys.argv: + with open(sys.argv[sys.argv.index("--json") + 1], "w") as file: + json.dump(rounded(out), file, indent=1) + file.write("\n") + + print("## Pause detection alone (no model)\n") + print("| Recording | Layout | Rule | Pieces | Under 2 s | Cut by length | Hand-overs caught | " + "Wait after the last word: median / p90 / max |") + print("| --- | --- | --- | --- | --- | --- | --- | --- |") + for c in out["cuts"]: + w = c["wait"] + print(f"| {c['recording']} | {c['layout']} | {c['rule']} | {c['pieces']} | {c['under2s']} | {c['forced']} | " + f"{c['caught']} of {c['handovers']} | {w['median']:.2f} / {w['p90']:.2f} / {w['max']:.2f} s |") + + print("\n## With a model\n") + print("| Model | Rec. | Layout | Pass | Pieces | Decode: median / max | Text after the pause: median / max | " + "Text after a hand-over's last word: median / p90 / max | Mean age | WER | CPU s per audio min | Load / first | Memory |") + print("| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |") + for p in out["passes"]: + d, t = p["decode"], p["delay"] + name = p["model"] + ("" if p["optimized"] else " (unoptimized)") + kind = "windows" if p["kind"] == "windows" else f"turns, {p['rule']}" + after = "" if p["afterPause"] is None else f"{p['afterPause']['median']:.2f} / {p['afterPause']['max']:.2f} s" + missing = f" ({p['missing']} never)" if p["missing"] else "" + cyr = f" ({p['cyrillic'] * 100:.0f}% Cyrillic)" if p["cyrillic"] > 0.01 else "" + recording = p["recording"] if p["language"] == p["recording"] else f"{p['recording']}, hint {p['language']}" + print(f"| {name} | {recording} | {p['layout']} | {kind} | {p['pieces']} | {d['median']:.2f} / {d['max']:.2f} s | " + f"{after} | {t['median']:.2f} / {t['p90']:.2f} / {t['max']:.2f} s{missing} | {p['age']:.1f} s | " + f"{p['wer']:.1f}%{cyr} | {p['cpuPerAudioMinute']:.1f} | {p['loadSeconds']:.1f} / {p['firstDecodeSeconds']:.2f} s | " + f"{p['footprintMB'] / 1024:.1f} GB |") + + if out["streams"]: + print() + print("## Streaming recognizer") + print() + print("| Model | Rec. | Chunk | Word shown after it is spoken: median / p90 / max | Text complete after a hand-over's " + "last word: median / p90 / max | Slowest 160 ms step | WER | CPU s per audio min | Load | Memory |") + print("| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |") + for r in out["streams"]: + a, d = r["age"], r["delay"] + cyr = f" ({r['cyrillic'] * 100:.0f}% Cyrillic)" if r["cyrillic"] > 0.01 else "" + print(f"| {r['model']} | {r['recording']}, hint {r['language']} | {r['chunkMilliseconds']} ms | " + f"{a['median']:.2f} / {a['p90']:.2f} / {a['max']:.2f} s | {d['median']:.2f} / {d['p90']:.2f} / {d['max']:.2f} s | " + f"{r['step']['max']:.2f} s | {r['wer']:.1f}%{cyr} | {r['cpuPerAudioMinute']:.1f} | {r['loadSeconds']:.1f} s | " + f"{r['footprintMB'] / 1024:.1f} GB |") + + if out["replays"]: + print() + print("## Through the app's live preview, in real time") + print() + print("| Model | Rec. | Played | Lines | Shown while speaking | Text after a hand-over's last word: median / p90 / max | " + "First words after a turn begins: median / max |") + print("| --- | --- | --- | --- | --- | --- | --- |") + for r in out["replays"]: + d, f = r["delay"], r["firstWords"] + missing = f" ({r['missing']} never)" if r["missing"] else "" + print(f"| {r['model']} | {r['recording']} | {r['seconds']:.0f} s | {r['lines']} | {r['partials']} | " + f"{d['median']:.2f} / {d['p90']:.2f} / {d['max']:.2f} s{missing} | {f['median']:.2f} / {f['max']:.2f} s |") + + +if __name__ == "__main__": + main() \ No newline at end of file diff --git a/Benchmarks/LiveTurns/split_tracks.py b/Benchmarks/LiveTurns/split_tracks.py new file mode 100644 index 00000000..18238aae --- /dev/null +++ b/Benchmarks/LiveTurns/split_tracks.py @@ -0,0 +1,50 @@ +#!/usr/bin/env python3 +"""Split a sample meeting's mixed recording into one track per speaker, the way a call is +recorded (microphone and call audio apart). Each moment goes to whoever is speaking, changing +hands at the middle of the pause between two speakers; the other track is digital silence there. +Needs ffmpeg and numpy. + +Usage: split_tracks.py +Writes ..wav for each speaker, in the order they first speak. +""" +import json +import subprocess +import sys +import wave + +import numpy as np + +RATE = 16000 + + +def main(): + recording, reference, prefix = sys.argv[1:4] + raw = subprocess.run( + ["ffmpeg", "-v", "error", "-i", recording, "-ac", "1", "-ar", str(RATE), "-f", "f32le", "-"], + check=True, capture_output=True).stdout + samples = np.frombuffer(raw, dtype=" 0 else { return nil } + quietChunks = chunks + eagerAfterSeconds = parts.count > 1 ? Double(parts[1]) : nil + } + + init(_ settings: LiveTurnCutter.Settings) { + quietChunks = settings.quietChunksToClose + eagerAfterSeconds = settings.eagerAfterSeconds + } + + var settings: LiveTurnCutter.Settings { + var settings = LiveTurnCutter.Settings() + settings.quietChunksToClose = quietChunks + settings.eagerAfterSeconds = eagerAfterSeconds + return settings + } + + /// Safe in a file name on every system. + var label: String { "\(quietChunks)q" + (eagerAfterSeconds.map { "-\(String(format: "%g", $0))s" } ?? "") } + } + + struct Cut: Codable { + var recording: String + var layout: String + var rule: Rule + var vadSecondsPerChunk: Double + var pieces: [Piece] + } + + struct Pass: Codable { + var layout: String + var kind: String + var rule: Rule? + var pieces: [Piece] + } + + struct Run: Codable { + var model: String + var modelName: String + var recording: String + var language: String + var audioSeconds: Double + var loadSeconds: Double + var firstDecodeSeconds: Double + var footprintMegabytes: Double + var optimized: Bool + var passes: [Pass] + } + + struct Stream: Codable { + var model: String + var recording: String + var language: String + var layout: String + var audioSeconds: Double + var loadSeconds: Double + var chunkMilliseconds: Int + /// Time to take in 160 ms of audio; most steps only buffer, one in every few runs the model. + var stepSeconds: [Double] + /// Seconds from the end of a word to its text existing, one entry per token. + var tokenAges: [Double] + var footprintMegabytes: Double + var optimized: Bool + var pieces: [Piece] + } + + struct Replay: Codable { + struct Arrival: Codable { + /// Seconds into the replay at which the preview showed this. + var at: Double + /// Where the line starts in the recording. + var time: Double + var speaker: String + var text: String + /// A turn still being spoken, not a finished line. + var partial: Bool + } + + var model: String + var recording: String + var language: String + var seconds: Double + var optimized: Bool + var arrivals: [Arrival] + } + + private struct Recording { + var label: String + /// "mixed" and, when the per-voice files were given, "tracks". + var layouts: [(name: String, tracks: [[Float]])] + var language: String { label.split(separator: "-").first.map(String.init) ?? label } + } + + private static let rate = Double(VadManager.sampleRate) + /// Pause rules compared without a model, beside the app's own. + private static let studied = ["1:off", "2:off", "3:off", "4:off", "2:3", "2:4", "2:6"].compactMap { Rule($0[...]) } + private static let englishOnly: Set = [.parakeetV2, .graniteTurbo, .graniteSpeech] + private static let candidates: [TranscriptionModelChoice] = [ + .parakeetV3, .parakeetUltra, .parakeetV2, .qwenASR17B, .qwenASR06B, + .whisperLarge, .graniteTurbo, .graniteSpeech, + ] + + func testTurnByTurnDelay() async throws { + let environment = ProcessInfo.processInfo.environment + guard let outputPath = environment["LOKALBOT_LIVE_TURN_BENCH"], + let audio = environment["LOKALBOT_LIVE_TURN_AUDIO"] else { + throw XCTSkip("Set LOKALBOT_LIVE_TURN_BENCH and LOKALBOT_LIVE_TURN_AUDIO to run this measurement.") + } + let output = URL(fileURLWithPath: outputPath, isDirectory: true) + try FileManager.default.createDirectory(at: output, withIntermediateDirectories: true) + let scratch = output.appendingPathComponent("scratch-\(UUID().uuidString)", isDirectory: true) + try FileManager.default.createDirectory(at: scratch, withIntermediateDirectories: true) + defer { try? FileManager.default.removeItem(at: scratch) } + + let wanted = environment["LOKALBOT_LIVE_TURN_MODELS"].map { Set($0.split(separator: ",").map(String.init)) } + // The rules a model is timed with; the app's own unless others are named. + let rules = Self.rules(environment) + + let recordings = try Self.recordings(audio) + XCTAssertFalse(recordings.isEmpty) + guard let vad = try Self.voiceActivity() else { + throw XCTSkip("The voice-activity model is not downloaded.") + } + let hints = Self.hints(environment) + + // Pause detection alone, for every setting: no model involved. + var cuts: [String: [Piece]] = [:] + for recording in recordings { + for layout in recording.layouts { + for rule in Set(Self.studied + rules).sorted(by: { $0.label < $1.label }) { + let cut = try await Self.cut(layout.tracks, vad: vad, rule: rule) + cuts["\(recording.label)/\(layout.name)/\(rule.label)"] = cut.pieces + try Self.write(Cut(recording: recording.label, layout: layout.name, rule: rule, + vadSecondsPerChunk: cut.secondsPerChunk, pieces: cut.pieces), + to: output.appendingPathComponent("cut-\(recording.label)-\(layout.name)-\(rule.label).json")) + print("live-turn cut \(recording.label) \(layout.name) rule=\(rule.label) pieces=\(cut.pieces.count) " + + "vad=\(String(format: "%.2f", cut.secondsPerChunk * 1000)) ms/chunk") + } + } + } + + for choice in Self.candidates where wanted?.contains(choice.rawValue) ?? true { + guard TranscriptionModelStore.isDownloaded(choice) else { + print("live-turn skip \(choice.rawValue): not downloaded") + continue + } + let engine = choice.engine + let loadStart = ContinuousClock.now + try await engine.prepare(progress: nil) + let loadSeconds = Self.seconds(since: loadStart) + for recording in recordings { + if Self.englishOnly.contains(choice), recording.language != "en" { continue } + let hint = hints[recording.label] ?? recording.language + let language: String? = hint == "auto" ? nil : hint + guard let mixed = recording.layouts.first?.tracks.first else { continue } + // The first decode after a load pays for compilation and caches. + let warm = try await Self.decode(Array(mixed.prefix(Int(6 * Self.rate))), engine: engine, + language: language, scratch: scratch) + var passes: [Pass] = [] + for layout in recording.layouts { + for rule in rules { + guard var pieces = cuts["\(recording.label)/\(layout.name)/\(rule.label)"] else { continue } + var previousReady = 0.0 + for index in pieces.indices { + let piece = pieces[index] + let result = try await Self.decode( + Self.slice(layout.tracks[piece.track], piece.start, piece.end), + engine: engine, language: language, scratch: scratch) + previousReady = max(piece.decision, previousReady) + result.seconds + pieces[index].decodeSeconds = result.seconds + pieces[index].cpuSeconds = result.cpu + pieces[index].ready = previousReady + pieces[index].text = result.text + } + passes.append(Pass(layout: layout.name, kind: "turns", rule: rule, pieces: pieces)) + } + if layout.name == "tracks" { + passes.append(Pass(layout: layout.name, kind: "windows", rule: nil, + pieces: try await Self.windows(layout.tracks, engine: engine, + language: language, scratch: scratch))) + } + } + let run = Run(model: choice.rawValue, modelName: choice.displayName, recording: recording.label, + language: hint, audioSeconds: Double(mixed.count) / Self.rate, + loadSeconds: loadSeconds, firstDecodeSeconds: warm.seconds, + footprintMegabytes: Self.footprintMegabytes(), optimized: Self.isOptimized, passes: passes) + let suffix = hint == recording.language ? "" : "@\(hint)" + try Self.write(run, to: output.appendingPathComponent("run-\(choice.rawValue)-\(recording.label)\(suffix).json")) + for pass in passes { + let decodes = pass.pieces.map(\.decodeSeconds).sorted() + guard !decodes.isEmpty else { continue } + print("live-turn run \(choice.rawValue) \(recording.label) \(pass.layout) \(pass.kind) " + + "rule=\(pass.rule?.label ?? "-") pieces=\(decodes.count) " + + "decode median=\(String(format: "%.2f", decodes[decodes.count / 2])) s " + + "max=\(String(format: "%.2f", decodes[decodes.count - 1])) s " + + "load=\(String(format: "%.1f", loadSeconds)) s first=\(String(format: "%.2f", warm.seconds)) s") + } + } + } + } + + /// The same replay through a streaming recognizer, which reads the audio + /// as it arrives and puts words out while the speaker is still talking. + /// The speech library the app links has one (Nemotron 3.5 streaming); the + /// app does not use or download it, so point this at a folder that holds + /// one variant's compiled models (Benchmarks/LiveTurns/fetch_stream_model.py): + /// + /// TEST_RUNNER_LOKALBOT_LIVE_STREAM_MODEL=/folder/multilingual-560ms \ + /// TEST_RUNNER_LOKALBOT_LIVE_TURN_LANGUAGES=en=en-US,sr=hr-HR + /// + /// Each turn is a fresh stream, opened a little before its first word and + /// closed when the pause after it is long enough. + func testStreamingRecognizerDelay() async throws { + let environment = ProcessInfo.processInfo.environment + guard let outputPath = environment["LOKALBOT_LIVE_TURN_BENCH"], + let audio = environment["LOKALBOT_LIVE_TURN_AUDIO"], + let modelPath = environment["LOKALBOT_LIVE_STREAM_MODEL"] else { + throw XCTSkip("Set LOKALBOT_LIVE_TURN_BENCH, LOKALBOT_LIVE_TURN_AUDIO and LOKALBOT_LIVE_STREAM_MODEL to run this measurement.") + } + let output = URL(fileURLWithPath: outputPath, isDirectory: true) + try FileManager.default.createDirectory(at: output, withIntermediateDirectories: true) + let rule = Self.rules(environment)[0] + let recordings = try Self.recordings(audio) + guard let vad = try Self.voiceActivity() else { + throw XCTSkip("The voice-activity model is not downloaded.") + } + let hints = Self.hints(environment) + let model = URL(fileURLWithPath: modelPath, isDirectory: true) + + let manager = StreamingNemotronMultilingualAsrManager() + let loadStart = ContinuousClock.now + try await manager.loadModels(from: model) + let loadSeconds = Self.seconds(since: loadStart) + let chunkMilliseconds = await manager.config.chunkMs + let step = Int(0.16 * Self.rate) + + for recording in recordings { + guard let layout = recording.layouts.last else { continue } + let hint = hints[recording.label] ?? recording.language + var pieces = try await Self.cut(layout.tracks, vad: vad, rule: rule).pieces + var ages: [Double] = [] + var steps: [Double] = [] + for index in pieces.indices { + let piece = pieces[index] + let samples = layout.tracks[piece.track] + await manager.reset() + await manager.setLanguage(hint == "auto" ? nil : hint) + let from = max(0, Int((piece.start - 0.3) * Self.rate)) + let to = min(samples.count, Int(piece.decision * Self.rate)) + let origin = Double(from) / Self.rate + var offset = from + var seen = 0 + let cpuStart = Self.cpuSeconds() + while offset < to { + let end = min(offset + step, to) + let started = ContinuousClock.now + _ = try await manager.process(samples: Array(samples[offset..() + var shown: [String: String] = [:] + var written = 0 + let total = min(layout.tracks[0].count, Int(seconds * Self.rate)) + let started = ContinuousClock.now + // Three more seconds after the last sample, for the last turn's text. + while Self.seconds(since: started) < Double(total) / Self.rate + 3 { + let due = min(total, Int(Self.seconds(since: started) * Self.rate)) + if due > written { + for (file, track) in zip(files, layout.tracks) { + try Self.append(Array(track[written.. AVAudioFile { + let settings: [String: Any] = [ + AVFormatIDKey: kAudioFormatLinearPCM, + AVSampleRateKey: rate, + AVNumberOfChannelsKey: 1, + AVLinearPCMBitDepthKey: 32, + AVLinearPCMIsFloatKey: true, + AVLinearPCMIsBigEndianKey: false, + AVLinearPCMIsNonInterleaved: true, + ] + return try AVAudioFile(forWriting: url, settings: settings, commonFormat: .pcmFormatFloat32, interleaved: false) + } + + private static func append(_ samples: [Float], to file: AVAudioFile) throws { + guard !samples.isEmpty, + let buffer = AVAudioPCMBuffer(pcmFormat: file.processingFormat, frameCapacity: AVAudioFrameCount(samples.count)), + let channel = buffer.floatChannelData?[0] else { return } + samples.withUnsafeBufferPointer { channel.update(from: $0.baseAddress!, count: samples.count) } + buffer.frameLength = AVAudioFrameCount(samples.count) + try file.write(from: buffer) + } + + // MARK: - Setup + + private static func recordings(_ audio: String) throws -> [Recording] { + var recordings: [Recording] = [] + for pair in audio.split(separator: ";") { + let parts = pair.split(separator: "=", maxSplits: 1) + guard parts.count == 2 else { continue } + // A second of quiet after the last word, so the last turn can close + // as it would while a recording goes on. + let files = try parts[1].split(separator: "|").map { + try AudioConverter().resampleAudioFile(URL(fileURLWithPath: String($0))) + + [Float](repeating: 0, count: Int(rate)) + } + guard let mixed = files.first else { continue } + var layouts = [(name: "mixed", tracks: [mixed])] + if files.count > 1 { layouts.append((name: "tracks", tracks: Array(files.dropFirst()))) } + recordings.append(Recording(label: String(parts[0]), layouts: layouts)) + } + return recordings + } + + /// The voice-activity model the app already has, or nil before its first download. + private static func voiceActivity() throws -> VadManager? { + let file = TranscriptionModelStore.Environment.live.fluidAudioModelsRoot + .appendingPathComponent(Repo.vad.folderName) + .appendingPathComponent(ModelNames.VAD.sileroVadFile) + guard FileManager.default.fileExists(atPath: file.path) else { return nil } + let configuration = MLModelConfiguration() + configuration.computeUnits = VadConfig.default.computeUnits + return VadManager(vadModel: try MLModel(contentsOf: file, configuration: configuration)) + } + + /// `LOKALBOT_LIVE_TURN_RULES=2:4,1:off` times those pause rules; without it, the app's own. + private static func rules(_ environment: [String: String]) -> [Rule] { + let named = (environment["LOKALBOT_LIVE_TURN_RULES"] ?? "").split(separator: ",").compactMap { Rule($0) } + return named.isEmpty ? [Rule(LiveTurnCutter.Settings())] : named + } + + /// `sr=hr` hands the model another language code for a recording; `sr=auto` hands it none. + private static func hints(_ environment: [String: String]) -> [String: String] { + var hints: [String: String] = [:] + for pair in (environment["LOKALBOT_LIVE_TURN_LANGUAGES"] ?? "").split(separator: ",") { + let parts = pair.split(separator: "=", maxSplits: 1) + if parts.count == 2 { hints[String(parts[0])] = String(parts[1]) } + } + return hints + } + + // MARK: - Turns + + /// Streams each track through the voice-activity model one 256 ms chunk + /// at a time, as a live feed arrives, and lets the app's `LiveTurnCutter` + /// close the turns. A turn that runs to the maximum length is cut at the + /// quietest point of its last 1.5 s, as the preview cuts its windows. + /// Pieces come back in the order they close. + private static func cut(_ tracks: [[Float]], vad: VadManager, + rule: Rule) async throws -> (pieces: [Piece], secondsPerChunk: Double) { + var pieces: [Piece] = [] + var chunks = 0 + let started = ContinuousClock.now + for (track, samples) in tracks.enumerated() { + var state = await vad.makeStreamState() + var cutter = LiveTurnCutter(settings: rule.settings) + var offset = 0 + while offset + VadManager.chunkSize <= samples.count { + let end = offset + VadManager.chunkSize + let result = try await vad.processStreamingChunk(Array(samples[offset.. [Piece] { + var pieces: [Piece] = [] + var clock = 0.0 + var processed = [Int64](repeating: 0, count: tracks.count) + while true { + var didWork = false + var waiting = false + for (track, samples) in tracks.enumerated() { + let total = Int64(samples.count) + let available = min(total, Int64(clock * rate)) + if available < total { waiting = true } + guard let range = LiveTranscriptChunker.nextChunk( + processedFrames: processed[track], totalFrames: available, sampleRate: rate) else { continue } + var chunk = Array(samples[Int(range.lowerBound).. 0 { + let offset = LiveTranscriptChunker.refinedCutOffset( + tail: Array(chunk.suffix(searchCount)), sampleRate: rate) + chunk.removeLast(searchCount - offset) + } + guard !chunk.isEmpty else { continue } + let chunkEnd = range.lowerBound + Int64(chunk.count) + didWork = true + processed[track] = chunkEnd + guard let candidate = LiveTranscriptChunker.speechCandidate(from: chunk, sampleRate: rate) else { continue } + let wallStart = ContinuousClock.now + let cpuStart = cpuSeconds() + let file = scratch.appendingPathComponent("window-\(UUID().uuidString).caf") + defer { try? FileManager.default.removeItem(at: file) } + try write(candidate, to: file) + var text = "" + if (await SpeechActivity.shared.speechSeconds(in: file) ?? 1) > 0 { + let transcript = try await engine.transcribe(audio: file, language: language, prompt: nil) + text = transcript.segments.map(\.displayText).filter { !$0.isEmpty }.joined(separator: " ") + } + let decode = seconds(since: wallStart) + pieces.append(Piece(track: track, start: Double(range.lowerBound) / rate, end: Double(chunkEnd) / rate, + decision: clock, forced: false, decodeSeconds: decode, + cpuSeconds: cpuSeconds() - cpuStart, ready: clock + decode, text: text)) + clock += decode + } + if !didWork { + // The last few seconds of a track wait for the recording to stop. + guard waiting else { break } + clock += 1 + } + } + return pieces + } + + // MARK: - Helpers + + private static func slice(_ samples: [Float], _ start: Double, _ end: Double) -> [Float] { + let lower = max(0, min(samples.count, Int(start * rate))) + let upper = max(lower, min(samples.count, Int(end * rate))) + return Array(samples[lower.. (text: String, seconds: Double, cpu: Double) { + let file = scratch.appendingPathComponent("turn-\(UUID().uuidString).caf") + defer { try? FileManager.default.removeItem(at: file) } + let wallStart = ContinuousClock.now + let cpuStart = cpuSeconds() + try write(samples, to: file) + let transcript = try await engine.transcribe(audio: file, language: language, prompt: nil) + let text = transcript.segments.map(\.displayText).filter { !$0.isEmpty }.joined(separator: " ") + return (text, seconds(since: wallStart), cpuSeconds() - cpuStart) + } + + /// 16 kHz mono float, the format the preview hands a model. + private static func write(_ samples: [Float], to url: URL) throws { + guard !samples.isEmpty else { throw CocoaError(.fileWriteUnknown) } + try append(samples, to: try liveFile(at: url)) + } + + private static func write(_ value: Value, to url: URL) throws { + let encoder = JSONEncoder() + encoder.outputFormatting = [.prettyPrinted, .sortedKeys] + try encoder.encode(value).write(to: url, options: .atomic) + } + + private static func seconds(since start: ContinuousClock.Instant) -> Double { + let elapsed = (ContinuousClock.now - start).components + return Double(elapsed.seconds) + Double(elapsed.attoseconds) / 1e18 + } + + /// CPU time this process has used, all threads. A model that runs in a + /// helper process (Granite Speech) is not counted. + private static func cpuSeconds() -> Double { + var usage = rusage() + getrusage(RUSAGE_SELF, &usage) + return Double(usage.ru_utime.tv_sec) + Double(usage.ru_utime.tv_usec) / 1e6 + + Double(usage.ru_stime.tv_sec) + Double(usage.ru_stime.tv_usec) / 1e6 + } + + private static func footprintMegabytes() -> Double { + var info = task_vm_info_data_t() + var count = mach_msg_type_number_t(MemoryLayout.size / MemoryLayout.size) + let result = withUnsafeMutablePointer(to: &info) { pointer in + pointer.withMemoryRebound(to: integer_t.self, capacity: Int(count)) { + task_info(mach_task_self_, task_flavor_t(TASK_VM_INFO), $0, &count) + } + } + return result == KERN_SUCCESS ? Double(info.phys_footprint) / 1_048_576 : 0 + } + + /// `assert` is compiled out of optimized builds, whatever flags are set. + private static var isOptimized: Bool { + var unoptimized = false + assert({ unoptimized = true; return true }()) + return !unoptimized + } +} diff --git a/Scripts/ci/unit-skips.json b/Scripts/ci/unit-skips.json index 5b79e405..5e257117 100644 --- a/Scripts/ci/unit-skips.json +++ b/Scripts/ci/unit-skips.json @@ -58,6 +58,14 @@ "CotypingTypingLatencyBenchmarkTests/testTypingLatency" ] }, + { + "reason": "Live transcript timing over downloaded speech models and recordings (LOKALBOT_LIVE_TURN_BENCH and LOKALBOT_LIVE_TURN_AUDIO, with LOKALBOT_LIVE_STREAM_MODEL for the streaming recognizer and LOKALBOT_LIVE_REPLAY_SECONDS for the real-time replay; Benchmarks/LiveTurns). Measurements, run by hand.", + "tests": [ + "LiveTurnLatencyBenchmarkTests/testPreviewReplayInRealTime", + "LiveTurnLatencyBenchmarkTests/testStreamingRecognizerDelay", + "LiveTurnLatencyBenchmarkTests/testTurnByTurnDelay" + ] + }, { "reason": "Needs cached Qwen3-ASR weights (LOKALBOT_QWEN_TEST_MODEL_DIR, and LOKALBOT_QWEN_TEST_REFERENCE for upgrade parity). Run for a speech runtime update.", "tests": [