A frameless desktop overlay that transcribes live system audio and translates it in real time — designed for watching foreign-language streams, videos, or calls without subtitles.
Audio is captured from the system's default output (loopback), transcribed by whisper.cpp, translated by a locally-running Ollama model, and displayed as streaming subtitles in an always-on-top overlay window.
- Continuous capture — audio is captured without gaps; transcription and translation run while the next chunk is already being buffered
- Streaming transcription — partial whisper segments appear word-by-word as the model decodes
- Streaming translation — Ollama tokens stream into the overlay in real time
- GPU acceleration — Vulkan (default) or CUDA; the active backend is shown at startup
- Always on top — stays visible over fullscreen windows
- Inline language selector — change source/target language from the status bar; saved automatically
The overlay displays source text (Japanese) in the top panel, live translation (English) below. Drag the header to reposition; use controls to toggle source view, access settings, or stop transcription.
| Dependency | Purpose |
|---|---|
| Ollama | Local LLM for translation |
| WebKit2GTK 4.1 | Web renderer (part of GNOME/KDE stack, usually pre-installed) |
| Vulkan loader | GPU acceleration |
| PipeWire or PulseAudio | System audio capture |
The AppImage bundles most dependencies. WebKit2GTK and GStreamer must come from the system.
Models are stored in ~/.cache/yaku/models/ and are not bundled.
| Model | Size | Notes |
|---|---|---|
ggml-tiny.bin |
75 MB | Fastest, lowest accuracy |
ggml-base.bin |
142 MB | Good for quick testing |
ggml-small.bin |
466 MB | Balanced |
ggml-medium.bin |
1.5 GB | Good accuracy |
ggml-large-v3-turbo.bin |
874 MB | Fast and accurate (recommended) |
ggml-large-v3.bin |
3.1 GB | Best accuracy |
- Download the AppImage from the Releases page.
- Make it executable and run:
chmod +x yaku-*.AppImage ./yaku-*.AppImage
- On first launch the app shows a model selection screen — choose and download a model directly from the UI.
See Developer setup below.
On first launch the app shows a model selection screen. Pick a size and click Download — the model is fetched from HuggingFace and saved to ~/.cache/yaku/models/ automatically. If you already have a .bin file, click Use an existing model file to point directly to it.
TranslateGemma is purpose-built for translation and is recommended:
ollama pull translategemma:4b # 3.3 GB — fast, good quality
ollama pull translategemma:12b # 8 GB — higher qualityGeneral-purpose models also work (llama3.2, qwen2.5:7b, etc.) but produce lower translation quality.
| Control | Action |
|---|---|
SRC |
Toggle source-language panel (resizes window) |
⚙ |
Open settings |
✕ |
Close |
▶ Start / ■ Stop |
Start or stop the pipeline |
Drag the header bar to move the window.
Shows the pipeline state (Listening / Transcribing / Translating) and the active language pair. Click either language code to change it immediately — changes are saved automatically.
| Setting | Description |
|---|---|
| Source / target language | Language pair for transcription and translation |
| Ollama model | Model to use for translation |
| Ollama URL | Defaults to http://localhost:11434/api/generate |
Config is saved to the platform config directory (~/.config/yaku/ on Linux, ~/Library/Application Support/yaku/ on macOS, %AppData%\yaku\ on Windows).
Debian / Ubuntu:
sudo apt-get install cmake glslc libvulkan-dev libgomp-dev g++ pkg-config \
libgtk-3-dev libwebkit2gtk-4.1-dev golang-goFedora / RHEL:
sudo dnf install cmake glslc vulkan-loader-devel libgomp-devel gcc-c++ pkgconf \
gtk3-devel webkit2gtk4.1-devel golangInstall Task:
go install github.com/go-task/task/v3/cmd/task@latestThe Wails CLI is installed automatically the first time you run
taskortask dev.
git clone --recurse-submodules https://github.com/opdude/yaku
cd yaku
task # compiles whisper.cpp on first run (~2 min), then builds the appOutput lands in ./bin/.
| Command | Description |
|---|---|
task |
Build bin/yaku (Vulkan GPU) |
task dev |
Wails dev mode with live frontend reload |
GPU=cuda task |
Build with CUDA acceleration |
task appimage |
Build a self-contained AppImage |
task test |
Run unit tests |
task clean |
Remove build output |
task clean-all |
Remove build output and compiled whisper libs |
audio.Capturer.Stream() platform audio capture
│
▼ 500 ms increments, rolling 10 s window
silence gate (VAD)
│
▼ when ≥ ChunkSeconds of speech detected
transcribe.Transcriber.Transcribe() whisper.cpp via CGO
├─ segment callbacks → live source text
└─ final text
│
▼
translate.Stream() Ollama NDJSON streaming
└─ token callbacks → overlay events
Capture, transcription, and translation are pipelined: while Ollama translates chunk N, the audio buffer is already filling with chunk N+1.
cmd/yaku/ Wails entry point, Go↔JS bridge (App struct)
frontend/ HTML/CSS/JS overlay UI
internal/
audio/ Capturer interface, VAD, PCM conversion; platform implementations
config/ Config struct, YAML persistence
transcribe/ whisper.cpp CGO wrapper, GPU backend detection
translate/ Ollama streaming, context-aware prompt construction
third_party/whisper.cpp/ git submodule, compiled by `task whisper`
build/ AppImage scripts and whisper build output
Releases are created automatically when a v* tag is pushed:
git tag v1.2.3
git push origin v1.2.3The release workflow builds a Vulkan AppImage via GoReleaser and uploads it to the GitHub release.