Files
stt-server/NOTES.md

10 lines
729 B
Markdown

# Notes
## TranscriptionInfo — unused fields
`model.transcribe()` returns a `TranscriptionInfo` object as its second value. We currently use `language` and `language_probability`. Other available fields:
- **`all_language_probs`** — full ranked list of `(language, probability)` tuples for the segment. Useful for debugging misdetection — e.g. when the model hallucinates Sinhala on noise, this would show Sinhala at the top with a high probability. Could be included in transcript events or exposed as a diagnostic endpoint.
- **`duration`** — total audio duration fed to the model.
- **`duration_after_vad`** — speech duration according to Whisper's internal VAD (not meaningful since we pass `vad_filter=False`).