Mobile dictation through the user's own environment server

[pedrozimmermannsoares]

Phone dictation is where T3 Code has no alternative: a third-party dictation app can't type into the mobile composer. Today the mic depends on Apple's on-device model, so Android has no dictation (#13603), iPhones before iOS 26 have none, and non-English speech is transcribed with the English model (#12743, pt-BR confirmed in the comments).

Proposal: let the mobile composer send the finished recording to the environment server it is already connected to, and transcribe it there with a local model. No API keys, no cloud service, and it uses hardware the user already runs. That keeps to the direction in #15422 and doesn't revisit the web/desktop scope decision in #11278.

I'd ship it in three separately approved PRs:

1. Server transcription with one opt-in local model (Parakeet), plus a mobile setting "Transcribe on: this phone / computer". Nothing changes by default.
2. Optional transcript cleanup with an editable prompt, through the text generation provider already configured.
3. Choosing the transcription model.

#14882 (@UtkarshUsername) already implements much of the server side for web/desktop. I'd reuse its approach for slice 1, with the author's agreement.

Is slice 1 in scope? If yes, I'll measure upload, transcription and cleanup latency on a real phone and post the numbers before opening the PR.

[Morriz]

+1, and a suggestion that might fold three open requests into one small surface.

We run a small team on T3 and dictate a lot from iPhone and iPad. The on-device engine struggles with long technical sentences: words drop out and sentences arrive half-garbled, even with a good mic in a quiet room. In our own agent platform we transcribe the same speech on the computer the phone talks to (Parakeet locally, an OpenAI-compatible service as fallback), and the difference is night and day.

What would cover this discussion, #28150 (choose the engine, BYOK) and #28636 (self-hosted STT endpoint) at once:

• One engine interface: an OpenAI-compatible transcription endpoint (POST /v1/audio/transcriptions) with a base URL, model and optional API key, configured on the environment server. That single client reaches OpenAI, Groq and any local server — Parakeet, Whisper and others already ship servers that speak this API — with no model bundling and no per-vendor code in T3.
• Transcribe through the environment server, as proposed here: the phone sends the finished recording to the server it is already connected to; the server calls the configured endpoint. Keys never live on the phone, Android and older iPhones get dictation, and the record → review → edit flow stays as it is.
• On-device stays the default; a setting "Transcribe on: this device / environment server" opts in.

Happy to help test, and to contribute if the direction fits.