[Bug]: Quick voice recording silently drops the audio when local STT is disabled (cloud transcription is never called)

[thiese]

Summary

Environment

• Version: 1.0.39+124 (main @ 4401c41)
• Platform: Android
• Provider: Gemini (API key), audio-capable model
• Settings > Speech transcription > "Use local speech to text": OFF

Summary

With local STT disabled, the quick recording (long-press on the AI core button) produces no transcript at all. The recorded audio is written to a temp WAV and then discarded without ever being sent to the configured LLM. The user only sees the "no speech detected" toast.

The cloud transcription code itself looks complete and correct — it is simply never invoked from anywhere in the app.

Steps to reproduce

1. Configure a provider whose model accepts audio input (e.g. Gemini Flash).
2. Settings > Speech transcription > turn OFF "Use local speech to text".
3. Long-press the center AI button, speak, release on an action.

Expected: the recording is transcribed by the configured model and opens the entry dialog as draft text.

Actual: no text appears during or after recording; a "no speech detected" toast is shown and the WAV is dropped.

Root cause

In lib/main.dart, _stopRecording() branches on isUsingLocalModel():

• local branch: calls SpeechTranscriptionService.transcribeSamples(samples) and assigns the result to _quickTranscribedText (~line 1362).
• cloud branch: saves the PCM buffer as a WAV and only stores the path in _quickAudioPath (~line 1375). No transcription call.

_quickTranscribedText can never be populated in cloud mode, because streaming transcription is local-only (supportsStreamingTranscription() returns false when not using the local model). So in _handleShortcutSelect() the empty-transcript branch is always taken and only shows speechNoResult (~line 1405), after which _quickAudioPath is set back to null and the file is never used.

Grepping the repo, transcribeFile, transcribeFileWithMetadata, transcribeSamplesWithMetadata and the private _transcribeFileWithCloud / _transcribeSamplesWithCloud in lib/data/services/speech_transcription_service.dart have no callers outside that file. The cloud path is effectively dead code.

Suggested fix

In the cloud branch of _stopRecording(), await SpeechTranscriptionService.instance.transcribeFile(wavPath) (or transcribeSamples(samples), which routes to the cloud implementation automatically) and assign the result to _quickTranscribedText, mirroring the local branch.

One thing to watch if the file-based variant is used: AssetSafetyService._inspectAudio() derives the duration from a numeric suffix in the file name. The quick recording is named quick_audio_<epoch_ms>.wav, so the epoch timestamp would be parsed as the duration in seconds and exceed maxAudioSecondsForAutoTranscribe (300), making the asset fail the safety check. The samples-based variant avoids the inspection entirely.

Impact

This makes voice input unusable for any language the bundled SenseVoice model does not support (it covers zh/en/ja/ko/yue). German speech, for example, is transcribed into English-sounding nonsense in local mode, and disabling local STT yields nothing at all — so there is currently no working path for those users.

Steps to reproduce

1. Configure a provider whose model accepts audio input (e.g. Gemini Flash).
2. Settings > Speech transcription > turn OFF "Use local speech to text".
3. Long-press the center AI button, speak, release on an action.

Expected behavior

Expected: the recording is transcribed by the configured model and opens the entry dialog as draft text.

Actual behavior

Actual: no text appears during or after recording; a "no speech detected" toast is shown and the WAV is dropped.

Platform

Android

Memex version or commit

No response

Logs, screenshots, or recordings

Privacy check

• I removed private records, API keys, screenshots with sensitive data, and local file paths that should not be public.