Recording captures no audio: all STT backends fail with SecurityException on FireOS, no audio file written, user sees silent 0-segment session #143

New issue

Closed

opened 2026-09-11 04:08:06 +00:00 by Mark · 0 comments

Mark commented

2026-09-11 04:08:06 +00:00

Owner

Bug

Tapping "Start recording" in the Recording screen appears to start a recording: the timer counts up, the Speaker/Subject chip rows appear, and tapping Stop creates a session in the Transcripts list with the recorded duration. But no audio is actually captured. The session row in the DB has audio_file_path = NULL and transcript_segments count is 0. The session detail shows "No transcript segments".

Reproduced on a Fire HD KFTRWI (Amazon FireOS), app com.voicetranscriber v1.0 built from 93b7fd1 (Stop silent STT credential fallback, 2026-09-10). Tested with a 2m 10s recording — UI behaved correctly, DB row existed, audio never written.

Root cause from logcat

SpeechBackend: Selected: ANDROID | available: [ANDROID, FIREOS] | missing: {CLOUD=[No API key configured], LOCAL_WHISPER=[Whisper model not found at /data/user/0/com.voicetranscriber/files/models/ggml-tiny.bin]}
TranscriptionService: Available backends: [ANDROID, FIREOS]
TranscriptionService: Attempting backend: ANDROID
TranscriptionService: Backend ANDROID failed (SecurityException): Not allowed to bind to service Intent { act=android.speech.RecognitionService cmp=amazon.speech.sim/com.amazon.alexa.multimodal.assistant.AlexaVoiceInteractionService }
TranscriptionService: Attempting backend: FIREOS
TranscriptionService: Backend FIREOS failed (SecurityException): Not allowed to bind to service Intent { act=android.speech.RecognitionService cmp=amazon.speech.sim/com.amazon.alexa.multimodal.assistant.AlexaVoiceInteractionService }
TranscriptionService: All backends failed to start. Stopping service.

Both ANDROID and FIREOS backends fail with SecurityException on the Fire HD because the system SpeechRecognizer routes to amazon.speech.sim/com.amazon.alexa.multimodal.assistant.AlexaVoiceInteractionService and the app cannot bind to it. LOCAL_WHISPER would work, but no model is downloaded (Models screen lists Tiny / Tiny (English) / Base as available, but BackendConfigScreen does not expose LOCAL_WHISPER in the priority list — see companion bug).

What the user sees

Tap Start recording → timer runs, mic indicator may or may not appear
Stop → session row appears in Transcripts with correct duration
Tap session row → "No transcript segments" message
No audio file is written to /data/data/com.voicetranscriber/files/ or external storage
No error message is shown to the user; the silent failure is not visible

Impact

The entire "Transcripts" view is non-functional on FireOS until a Whisper model is downloaded AND LOCAL_WHISPER is selected as the active backend
The app gives no indication that the recording failed — users assume their voice was captured
The Speaker/Subject chip selection during recording is lost because the session has no segments to attribute
All downstream features (Notes, Search, Diarization, Topic Extraction, Summarization, Session Tags) are unreachable because they all gate on segments.isNotEmpty()

Suggested fix directions

Surface a visible error when all backends fail: the Recording screen should show "Recording failed — no transcription backend available. Download a Whisper model or check your microphone permissions." instead of silently showing a 0-segment session.
Auto-record the raw audio via MediaRecorder even if the SpeechRecognizer backend fails, so users at least get a .m4a file that can be transcribed later.
Fix the LOCAL_WHISPER BackendConfigScreen bug (companion issue) so users can promote LOCAL_WHISPER to the top of the priority list before the ANDROID/FIREOS backends.
Fix the FireOS SpeechRecognizer binding — either request a special permission for amazon.speech.sim/... on FireOS devices, or detect FireOS and skip the ANDROID/FIREOS backends entirely when LOCAL_WHISPER is the only realistic option.

Evidence

DB before fix: (1, 'Untitled recording', 1789098115153, 1789098127684, None, None) — 12s test, no audio
DB after fix: (2, 'Untitled recording', 1789098766618, 1789098897057, None, None) — 2m 10s test, no audio
Logcat timestamps from the 2m 10s attempt: 20:52:46.755 → 20:52:46.759 — both ANDROID and FIREOS fail in 4ms
Models screen DOES list Whisper models for download, but LOCAL_WHISPER card is missing from the BackendConfigScreen priority list (companion issue)

Reported by Hermes on 2026-09-10 during smoke testing of v1.0 (build from 93b7fd1) on Fire HD KFTRWI (G001KT05122500T3).

## Bug Tapping "Start recording" in the Recording screen appears to start a recording: the timer counts up, the Speaker/Subject chip rows appear, and tapping Stop creates a session in the Transcripts list with the recorded duration. **But no audio is actually captured.** The session row in the DB has `audio_file_path = NULL` and `transcript_segments` count is 0. The session detail shows "No transcript segments". Reproduced on a Fire HD KFTRWI (Amazon FireOS), app `com.voicetranscriber` v1.0 built from `93b7fd1` (Stop silent STT credential fallback, 2026-09-10). Tested with a 2m 10s recording — UI behaved correctly, DB row existed, audio never written. ## Root cause from logcat ``` SpeechBackend: Selected: ANDROID | available: [ANDROID, FIREOS] | missing: {CLOUD=[No API key configured], LOCAL_WHISPER=[Whisper model not found at /data/user/0/com.voicetranscriber/files/models/ggml-tiny.bin]} TranscriptionService: Available backends: [ANDROID, FIREOS] TranscriptionService: Attempting backend: ANDROID TranscriptionService: Backend ANDROID failed (SecurityException): Not allowed to bind to service Intent { act=android.speech.RecognitionService cmp=amazon.speech.sim/com.amazon.alexa.multimodal.assistant.AlexaVoiceInteractionService } TranscriptionService: Attempting backend: FIREOS TranscriptionService: Backend FIREOS failed (SecurityException): Not allowed to bind to service Intent { act=android.speech.RecognitionService cmp=amazon.speech.sim/com.amazon.alexa.multimodal.assistant.AlexaVoiceInteractionService } TranscriptionService: All backends failed to start. Stopping service. ``` Both ANDROID and FIREOS backends fail with `SecurityException` on the Fire HD because the system SpeechRecognizer routes to `amazon.speech.sim/com.amazon.alexa.multimodal.assistant.AlexaVoiceInteractionService` and the app cannot bind to it. LOCAL_WHISPER would work, but no model is downloaded (Models screen lists Tiny / Tiny (English) / Base as available, but BackendConfigScreen does not expose LOCAL_WHISPER in the priority list — see companion bug). ## What the user sees 1. Tap Start recording → timer runs, mic indicator may or may not appear 2. Stop → session row appears in Transcripts with correct duration 3. Tap session row → "No transcript segments" message 4. No audio file is written to `/data/data/com.voicetranscriber/files/` or external storage 5. No error message is shown to the user; the silent failure is not visible ## Impact - The entire "Transcripts" view is non-functional on FireOS until a Whisper model is downloaded AND LOCAL_WHISPER is selected as the active backend - The app gives no indication that the recording failed — users assume their voice was captured - The Speaker/Subject chip selection during recording is lost because the session has no segments to attribute - All downstream features (Notes, Search, Diarization, Topic Extraction, Summarization, Session Tags) are unreachable because they all gate on `segments.isNotEmpty()` ## Suggested fix directions 1. **Surface a visible error** when all backends fail: the Recording screen should show "Recording failed — no transcription backend available. Download a Whisper model or check your microphone permissions." instead of silently showing a 0-segment session. 2. **Auto-record the raw audio** via `MediaRecorder` even if the SpeechRecognizer backend fails, so users at least get a `.m4a` file that can be transcribed later. 3. **Fix the LOCAL_WHISPER BackendConfigScreen bug** (companion issue) so users can promote LOCAL_WHISPER to the top of the priority list before the ANDROID/FIREOS backends. 4. **Fix the FireOS SpeechRecognizer binding** — either request a special permission for `amazon.speech.sim/...` on FireOS devices, or detect FireOS and skip the ANDROID/FIREOS backends entirely when LOCAL_WHISPER is the only realistic option. ## Evidence - DB before fix: `(1, 'Untitled recording', 1789098115153, 1789098127684, None, None)` — 12s test, no audio - DB after fix: `(2, 'Untitled recording', 1789098766618, 1789098897057, None, None)` — 2m 10s test, no audio - Logcat timestamps from the 2m 10s attempt: `20:52:46.755` → `20:52:46.759` — both ANDROID and FIREOS fail in 4ms - Models screen DOES list Whisper models for download, but LOCAL_WHISPER card is missing from the BackendConfigScreen priority list (companion issue) Reported by Hermes on 2026-09-10 during smoke testing of v1.0 (build from 93b7fd1) on Fire HD KFTRWI (G001KT05122500T3).