[Feature]: Chinese (zh-CN) speech recognition and desktop composer voice

[sprintthunder-dotcom]

Problem

Voice dictation in T3 Code is iPhone-only today, and the shipped iOS path does not treat Chinese (or mixed Chinese–English) as a first-class locale.

I write prompts in Chinese and in mixed Chinese–English (identifiers, paths, and commands inside a Chinese sentence). On iPhone the composer already has a microphone. On T3 Code desktop there is no equivalent control, so the same workflow is keyboard-only.

This is not UI localization. It is speech-to-text into the composer draft.

Current status (verified 2026-09-07)

• Shipped: on-device voice input on supported iPhones (iOS 26+). Docs: Voice input on iPhone. Internals: locale is bound once per recording; environment-backed transcription is not implemented. Merged in feat(mobile): add offline iPhone voice input #8614.
• Desktop: not shipped. Maintainer-owned local desktop implementation feat(desktop): add local voice input #8010 was closed (2026-08-28) because capture + local STT + model download + IPC needed a clearer native/privacy/support plan.
• Open implementation PRs (not asking you to merge these as-is): feat: add voice dictation beta #5213 (web/desktop BYOK dictation beta), feat(web): dictation beta with mic button, cleanup and shortcut #10195 (web dictation beta, includes a language setting), [WIP] feat(voice): transcribe on the environment #8928 (WIP: transcribe on the environment so desktop stays local and web can share it).
• Closed related asks: [Discussion] Voice support #653 (voice discussion), [Feature]: Mobile voice dictation — mic button in the composer with BYOK OpenAI/Groq transcription #8718 (mobile BYOK dictation, not planned), [Feature]: T3 never requests macOS microphone/camera permission, so it can never be granted - blocks all voice/vision MCP tooling #7242 (macOS mic TCC for MCP tools — different from composer dictation).
• No existing Idea for Chinese STT or desktop composer voice. zh-CN discussions ([Feature]: Community-led Simplified Chinese localization with a minimal i18n foundation #6789, [Feature]: Simplified Chinese (zh-CN) localization for web, desktop and mobile apps #6900, We need Chineses！ #7384) are UI i18n, not speech.

Proposed solution

Two scoped product requirements, one capability:

1. Chinese speech recognition on whatever dictation path you ship

   • Support zh-CN (and mixed Chinese–English in one utterance).
   • On iOS, Apple's transcriber already binds a locale; make Chinese a supported/tested locale, not only the device language if that language is English.
   • Mixed input example: speaking 把 src/auth.ts 里的 verifyToken 改掉 should land as editable composer text with both Han characters and Latin identifiers, not as pinyin-only or English-only.
   • Do not auto-send. Insert at the caret, same as the current iPhone behaviour.

2. Desktop composer voice, parity with iPhone

   • Mic control on T3 Code desktop (macOS first is enough).
   • Record → transcribe → insert into the draft. No cloud requirement if the iOS path stays on-device; if you choose BYOK/environment transcription instead, desktop should use that same path so we do not grow a third engine.
   • Reuse the existing open PRs' direction if it matches this: environment-local ([WIP] feat(voice): transcribe on the environment #8928) or opt-in BYOK (feat: add voice dictation beta #5213 / feat(web): dictation beta with mic button, cleanup and shortcut #10195). I am not asking for a new native stack like the closed feat(desktop): add local voice input #8010 Moonshine download.

Smallest useful scope

• iOS: zh-CN + mixed Chinese–English on the existing Apple transcriber.
• Desktop: one mic button that inserts a transcript into the composer.

Out of scope for this ask: TTS, sending raw audio to the agent, Android, UI translation.

Why this matters

Chinese-speaking developers use T3 from the phone and the desktop. Typing mixed Chinese–English is slow. Voice is already in the iPhone composer; without Chinese STT and without desktop, the feature is unusable for that workflow.

Contribution

I can test iOS + macOS desktop with a Chinese IME and mixed utterances, and report transcripts. I am not proposing a specific STT vendor.

[olafura]

This could be easilly done with #10708 where you could just hook up whatever input capability you wanted

[olafura]

This gist is a demonstration of how this feature could be built or customized with the Qt/QML branch. It is an example to try and adapt, not a complete implementation of the discussion request or a feature shipped in upstream T3 Code.

I added an opt-in local dictation example for the Qt composer. Once configured with an existing Python environment and faster-whisper model, it uses Linux PipeWire capture and local CPU transcription, then inserts an editable transcript at the caret without sending it. zh-CN selects Whisper's Chinese transcription mode.

It installs or downloads nothing and stays disabled until configured. A result arriving after a thread switch is retained for copying instead of being inserted into the wrong draft. The gist also includes a paste-transcript alternative.

The actual Dictate and Stop controls were tested with prerecorded Mandarin audio and an existing local faster-whisper large-v3 model. The recognizer inserted 一二三四五六七八九 into the composer without sending it. No microphone was used. This narrow fixture check does not establish live microphone quality or mixed Chinese-English recognition accuracy. This doesn't implement macOS capture or the requested iOS locale support, so it isn't the complete feature request.

Demonstration gist and setup.

The screenshot shows that real recognition result on workspace 5, alongside demo conversation data.

Code and checks assisted by Codex.