Transcription and audio
Languages, accuracy, fallback behavior, and what happens to your audio.
Reference for how KTB turns a consultation into text — and what it does with the audio afterwards.
Languages
Transcription is tuned for Portuguese (pt-BR) consults. Speaker turns are labeled automatically so the transcript reads as a conversation, not a wall of text.
Working in another language? Transcription quality is best in Portuguese — test a short session before relying on it for a full consult.
How transcription works
KTB uses a client-first strategy:
- Browser transcription runs in the browser as you speak — fast and private-by-default.
- Confidence check. After the session, KTB measures how much actual speech was captured versus transcript output.
- Server fallback. If confidence is suspiciously low (very little speech detected, or almost no transcript text), the recorded audio is sent for server-side re-transcription with a larger speech model.
You never have to choose — the fallback is automatic. When it triggers, the consultation keeps both versions: the raw server transcript and an LLM-cleaned transcript, clearly labeled as machine output pending your review.
Accuracy
- Expect a good first draft, not a perfect one — names, medications, and numbers are exactly what you should proofread.
- The silence/noise gate prevents the model from inventing phrases during quiet moments.
- Speaker labels can be wrong in overlapping speech; fix them in review.
What happens to the audio
- Session audio is stored encrypted-at-rest in Cloudflare R2, keyed to the consultation, so the server pipeline can re-transcribe when needed.
- Audio is never shown in error messages — pipeline failures report a reason, not the recording.
- The cleaned transcript is explicitly marked as machine-corrected and not physician-approved until you say so.
FAQ
Does transcription work offline? Browser transcription needs network for model assets on first use; the server fallback requires connectivity. Plan for online consults.
Can I delete the audio? Audio retention follows your clinic's data practices — see Privacy and consent and ask your administrator about retention policy.
Why does the transcript look different after processing? You are seeing the server-side re-transcription plus an LLM cleanup pass. Compare it with the browser transcript on the detail page and keep whichever reading is accurate — or edit directly.