[Feature]: Choose transcription engine (BYOK — Gemini/OpenAI/Groq) for macOS voice dictation

[abdullah-elbedwehy]

Request

The macOS app already has native voice dictation (shipped in #24938). This asks for one addition: let the user pick the transcription engine and bring their own API key, instead of only the OS/native engine. Gemini is my preferred engine, but the point is a provider choice, not one specific vendor.

Problem / use case

Native (OS) transcription quality varies a lot by language and accent, and it can't be tuned. For non-English and for technical/project vocabulary it mangles words often (same pain as the completed #27752 vocab-hints issue, but hints only go so far on the native engine). A cloud STT engine I choose — e.g. Gemini, OpenAI Whisper, Groq — is usually much more accurate and I already have keys for it. The dictation UX from #24938 (record, review, edit before send) stays exactly the same; only the engine that produces the transcript changes.

Related and distinct:

• [Feature]: Native voice transcription in the macOS app #24938 (Native voice transcription, macOS) — completed; this builds on it, swapping the engine.
• [Feature]: Mobile voice dictation — mic button in the composer with BYOK OpenAI/Groq transcription #8718 (mic + BYOK OpenAI/Groq) — mobile, closed not-planned. This one is desktop and provider-agnostic.
• Voice input: project-aware vocabulary hints for speech-to-text #27752 (vocab hints) — completed; helps the native engine, doesn't add external engines.

Desired behavior

• A Settings option under voice/transcription: Transcription engine = Native (OS) (default, current) or a BYOK cloud engine.
• For a BYOK engine: pick provider (e.g. Gemini / OpenAI / Groq), paste API key, optional model and language.
• The key is stored in the same encrypted secrets store the app already uses for provider credentials; never logged.
• Recording UX unchanged (start/stop shortcut, review/edit transcript, no auto-send, cancel, clear error on failure).
• On a transcription failure or missing/invalid key, fall back to the native engine (or show a clear error), so the mic never silently dies.

Motivation

Much better accuracy for non-English and technical dictation, using a key I already own, without changing the dictation flow I already use.

Platform

macOS desktop (T3 Code Beta 0.1.5). Same idea applies to other desktop platforms.

[LegendarySpy]

+1

This would be very helpful. As someone who ships a dictation app, it would be nice to reuse the existing engine and model files instead of having to duplicate and download more. A custom base URL for any OpenAI-compatible endpoint would cover that, plus Groq and local servers.