The speech-to-text function crashes the application on Windows.

[insign]

Whenever I try to use the speech-to-text function on Windows, the application closes automatically. This happens whether I use the native option or other models like Whisper or ParaKeet

[insign]

Re-investigation (v1.126.1 / 16fd7c0b)

Reproduction still holds on the current main for both engines the user reported. The "crash" claim deserves a split, because each engine has a different failure path on Windows today.

Two distinct failure paths, one user-visible symptom

1. Native (speech_to_text / UWP recognizer) — still the only available path on Windows.

   • lib/presentation/services/speech_input_service_stt.dart:59-78 calls SpeechToText.initialize() and on _isAvailable == false builds an actionable unavailableReason via WindowsMicrophoneService.probe() (lib/presentation/services/windows_microphone_service.dart:81-93).
   • No fallback / try-catch around the actual listen() call. startListening (speech_input_service_stt.dart:125-156) awaits speechToText.listen(...) with no exception guard. If the OS recognizer hands back a fatal COM error (e.g. E_FAIL from SpeechRecognizer.RecognizeAsync when no language pack is installed, or the runtime is missing entirely), the exception propagates out of the engine and into the chat-input speech controller.
   • chat_input_speech_controller.dart:239 and the surrounding startListening flow don't have a structured "engine exception" path either — they just rely on onError callbacks. A genuinely unhandled async error there is the most plausible source of a hard process exit on Windows.
   • ADR-039 partially addressed the friendly preflight (typed reason keys + Settings links), but the runtime still trusts the plugin to never throw.

2. On-device engines (Sherpa, Moonshine, Parakeet, SenseVoice) — intentionally disabled on Windows today.

   • lib/presentation/utils/speech_engine_platform_support.dart:31-65 returns false for all four on Windows.
   • lib/presentation/services/speech_audio_capture.dart:55-62 throws StateError("…not available on Windows; the on-device engines are disabled there until the WASAPI capture backend is validated (ADR-039).") when startPcmStream is called.
   • That StateError is exactly why the user sees "model downloads don't work" — the download itself succeeds (the model managers still write into getApplicationSupportDirectory()), but the engine cannot consume the downloaded files on Windows, so transcription never starts. There is no progress card explaining that the model is downloaded but unusable, which from the user's perspective looks like the download failed.

Why the WASAPI fix never landed

For historical context, this was almost fixed:

• 4a148061 fix: real Windows STT fix via WASAPI capture and actionable settings links implemented a full WindowsPcmCaptureService (PCM16 mono 16 kHz) and re-enabled all four on-device engines. It also added the typed preflight and the three Settings links the user sees today.
• 889a382a fix: address reviewer findings for issue #43 partial fix reverted the WASPI re-enable ("shipping a Windows WASAPI capture backend without a Windows CI gate risks reintroducing the very crash issue The speech-to-text function crashes the application on Windows. #43 reports"). The actionable links + preflight + lifecycle cleanup were kept as the partial fix; the new audio capture layer was removed.
• 0dd42d7d fix: tighten SpeechAudioCapture lifecycle and snackbar reason lookup further polished the partial fix.

Net result: settings links + typed preflight shipped, on-device engines did not.

What the partial fix actually does for the user right now

• The settings card lib/presentation/pages/settings/sections/speech_settings_section.dart:282-352 renders on Windows with three buttons (microphone privacy, speech privacy, speech settings) backed by WindowsSettingsLinks (the official ms-settings: URIs). For most users, this is the real fix for the "Native doesn't work" complaint — the underlying root cause is usually a missing language pack or a privacy toggle, both of which the buttons address.
• The chat-input snackbar (post-0dd42d7d) carries the same typed reason key and reuses the same Settings link.
• The download dialogs (sherpa_model_download_dialog, moonshine_model_download_dialog, parakeet_model_download_dialog, sensevoice_model_download_dialog) all look available on Windows because the engine toggle is just disabled — but the file picker / download still proceeds. That's the source of the second user complaint: "the download seemed to succeed but nothing happens".

Diagnostic questions to add to the issue

Before we can call this fixed again, we need the user to confirm which engine they actually clicked, and whether the app genuinely crashed vs. just failed silently:

1. Which engine was selected in Settings → Speech when the failure happened? (Native / Sherpa / Moonshine / Parakeet / SenseVoice)
2. App version? main SHA / About screen reports v1.126.1 (16fd7c0b) on the latest.
3. Crash vs. silent failure?
   • Crash: Windows Event Viewer → Windows Logs → Application, look for a .NET Runtime / .NET Framework / Application Error entry with CodeWalk.exe as faulting module. Paste the exception code and module name.
   • Silent: did the speech indicator appear in the composer, then disappear with no text? Did the Settings card "Disabled on Windows" subtitle appear for Moonshine/Parakeet/SenseVoice/Sherpa? (They should.)
4. Did the Windows setup card with the three Settings buttons appear under the engine list on Windows? (speechWindowsSetupHint — lib/l10n/.../app_localizations_en.dart around speechWindowsSetupHint.)
5. Does clicking each of the three Settings links open the correct Windows settings page?
6. After clicking Settings → speech (third button) and verifying "Online speech recognition" is on plus a language pack is installed, does the Native engine start working?

Path forward (proposed)

Three independent tracks. The first one is small and would unblock the on-device engines today; the second needs a Windows CI runner we already pin in .github/workflows; the third is the only way to actually close #43.

1. Short term (low risk). When the user picks an on-device engine on Windows, surface a card before download explaining "Models can be downloaded, but transcription is disabled on Windows until the WASAPI capture backend ships. Native STT is the only working option." That removes the "downloaded but useless" dead-end the user is hitting. File: speech_settings_section.dart:354-456 (the engine radio tiles).

2. Medium term. Re-enable the WASAPI capture backend in a feature branch with a Windows CI gate that:

   • Builds the native side on windows-2025 (we already pin it, see ci(release): pin Windows runner to windows-2025 fdba4b41).
   • Runs a smoke test that initializes WindowsPcmCaptureService and reads 1s of PCM16 (no decode, no model — just capture).
   • Without that gate, we cannot guarantee the crash won't return.

3. Long term / real close. Either (a) ship the validated WASAPI backend and re-apply the re-enable that was reverted in 889a382a, and wrap speechToText.listen (speech_input_service_stt.dart:142-155) in a try/catch that funnels COM exceptions into the typed preflight reason key, or (b) migrate to record: 7.x / record_windows: 2.x if the upstream MediaFoundation crash is finally fixed in a newer record line. Currently blocked by SDK constraint per ADR-039.

Acceptance criteria refresh

• Reproduction split confirmed (crash vs. silent no-op) with logs.
• User-facing card added so Windows users on a disabled engine know that "download" is purely local storage and the engine will not start.
• speechToText.listen on Windows is wrapped in a try/catch that maps the failure to the typed unavailableReasonKey.
• Windows CI gate exists for any future WASAPI re-enable.
• WASAPI backend re-applied and validated by that CI gate.
• record plugin upgrade path tracked (separate ticket) for users who want the upstream fix instead of the WASAPI shim.

References

• Code: lib/presentation/services/speech_audio_capture.dart, lib/presentation/services/speech_input_service_stt.dart, lib/presentation/utils/speech_engine_platform_support.dart, lib/presentation/services/windows_microphone_service.dart, lib/presentation/pages/settings/sections/speech_settings_section.dart.
• ADRs: ADR-038 (disable on-device engines on Windows), ADR-039 (partial fix / typed preflight / Settings links), ADR-006 (speech input architecture).
• Commits: 25ad52d0 (initial crash mitigation), 4a148061 (real fix, reverted), 889a382a (revert + reviewer pass), 0dd42d7d (lifecycle tighten).
• Upstream: llfbandit/record#453 (MediaFoundation segfault), record: ^6.0.0 / record_windows 1.0.7.