Faster live transcription
The real-time speech pipeline moved to a streaming decoder, so text appears sooner during recording and holds a steadier pace through a long session.
What's new
A streaming decoder for live transcription
Real-time speech recognition now runs on a streaming decoder, which shortens the delay between someone speaking and their words appearing in the transcript.
Sentence segmentation separated from final text
The live pass decides where sentences break; the higher-quality pass produces the wording that is kept. Splitting the two lets each be tuned for what it is actually doing.
Improvements
- The decoder is warmed before a session reports itself ready, so the first sentence of a recording is no longer the slowest.
- Real-time processing was brought back within the budget of the deployed GPU, which removes the drift that built up over long sessions.
- Voice activity detection was updated to match the deployed model, improving where speech is detected to start and stop.
Fixes
- Fixed a mismatch between the client and the deployed voice activity detection schema.
- Fixed pacing in long recordings where the transcript could fall progressively further behind the audio.