[Bug]: Mobile dictation has no usable German transcription #19091

Area

apps/mobile

Steps to reproduce

1. Set the iPhone device language to German (de-DE) with Apple Intelligence speech assets available.
2. Open any thread composer in the mobile app and tap the microphone.
3. Dictate a German sentence and confirm.

On Android: open any composer — no microphone action is offered at all.

Expected behavior

German dictation is transcribed in German, offline, on both iOS and Android.

Actual behavior

• iOS: transcription goes through Apple on-device SpeechTranscriber only (apps/mobile/src/native/voiceTranscription.ios.ts), gated on iOS 26+ and supportedLocale(equivalentTo:). With a German device locale there is no usable German path in practice — transcription fails (Voice transcription does not support this device language) or comes back English. Related: [Bug]: iPhone 1.2.0 voice input recognizes Korean speech as unrelated English words #12743 (Korean recognized as unrelated English words) — same underlying limitation for non-English locales.
• Android: getLocalVoiceTranscriber() (apps/mobile/src/native/voiceTranscription.ts) unconditionally returns null, so dictation is entirely unavailable.

Impact

Major degradation or frequent failure (German-language users; Android users are blocked completely).

Version or commit

main, October 2026 (voice controller with native-audio timeouts, cf. #19063).

Environment

iPhone, iOS 26, device language de-DE; Android (any version — no transcriber at all).

Suggested direction

Run transcription on the T3 server instead of only on the phone, using the exact stack I use daily for German dictation on desktop: winidi/dictate (MIT) with its local backend — parakeet-primeline, a German fine-tune of nvidia/parakeet-tdt-0.6b-v3, as int8 ONNX decoded CPU-only via sherpa-onnx (~50 ms decode for a 2 s clip at 4 threads, no network, no API key). The mobile app already abstracts transcription behind the VoiceTranscriber seam (prepare/transcribe, selected via getTranscriber in VoiceInputProvider), so a server-backed transcriber fits without touching the recording lifecycle:

• server side: sherpa-onnx-node (npm, node-addon-api, multithreaded) supports OfflineRecognizer transducer configs on macOS/Linux, i.e. the same pinned model files dictate uses (flozen1981/parakeet-primeline-onnx, ~640 MB, opt-in download on first use);
• client side: upload the finished recording to the server and await text — this also enables dictation on Android, where no on-device transcriber exists today;
• keep Apple on-device transcription as the fallback when the server has no model.

Notes: the model is CC-BY-4.0 (attribute primeline / NVIDIA / sherpa-onnx, as dictate does). I run dictate locally with small macOS usability improvements (hotkey guards for keys pynput lacks on macOS, Cmd-double-tap toggle); happy to share details if this direction is accepted. I did not open a code PR because this needs maintainer approval on scope (new server dependency + model download) first.

Workaround

Dictate externally on desktop; nothing in-app.

Muse Spark via OpenCode.