No text does not mean a broken model: capture, input, model and empty-result layers
The same missing-text symptom can come from different layers
The controller separates permission denial, recording/PCM failures and missing/load-model errors. ASR checks nonempty finite samples/length, and empty final text produces emptyTranscript. Short speech maps empty results to recording failure; long notes can skip text and advance frames. Notices are categories, not complete diagnosis or proof the original contained nothing.
Different from admission, integrity and retry checkpoints
Admission checks input, integrity checks assets/readiness, checkpoints check retry position. Here error origin and limits of visible categories matter. No capture/model/retry/file operation ran.
| Check | Verified scope | Manual review | Not established |
|---|---|---|---|
| Permission | captureFailure has microphonePermission case | Actual system state | Every error is permission |
| Recording/PCM | Vetted errors/recording fallback | Format/interruption/input | Model certainly broken |
| ASR array | Nonempty/finite/length checks | Array versus hardware format layers | Valid array proves meaningful speech |
| Model missing/load | Model notice category | Integrity/loading evidence | Only cause is hash tampering |
| emptyTranscript | Empty returned final | Actual audio/meaning/stage | Nobody spoke |
| Caller difference | Short speech/long notes differ | Failure versus skipped-piece progress | Every entry has identical outcomes |
| Notice copy | Arbitrary raw system errors not exposed | Category versus full diagnostics | All causes established |
| Next steps | Original authority/target/pending guards | Minimal nonsensitive scope | Automatic deletion/cloud/replay |
Synthetic layered note
emptyTranscript establishes no usable final from this pass, not a corrupt model or empty meeting. A model-load notice likewise needs actual integrity/runtime review instead of a tampering claim.
Advanced frames do not certify every word; notices are not service receipts. No device/real-speech regression ran. Do not publish recordings/system paths/account contents to investigate.
FAQ
Does no text always mean corrupt model?
No. Input, capture, permission, empty results and loading differ.
Does emptyTranscript prove silence?
Not independently. Actual audio/content need review.
Are empty-result callers identical?
No. Short failure and long-note skip/progress differ.
Does load failure prove malicious files?
No. It is not complete root-cause diagnosis.
Are notices full raw system errors?
No. Classification/fallback differs from full diagnostics.
Was a model deleted/recording/cloud fallback invoked?
No. No automatic deletion/fallback/replay is added; actual operations retain authority/guards.
Reading and source
Input format · Model evidence · Retry position · Permission layers · Start · Privacy
Error categories · Sample/final checks · Empty pieces/progress
