Starstruk Logo MarkStarstruk

0.2.6

QuentinFuxa/WhisperLiveKit0.2.6Aug 21, 2025by QuentinFuxa

AI Summary

This release (0.2.6) of WhisperLiveKit introduces significant default behavior changes including enabling Voice Activity Control (VAC) by default and making the simulstreaming backend the default transcription method, both aimed at improving transcription accuracy and resource management.

Key Highlights

  • Voice Activity Control (VAC) is now enabled by default to filter non-speech segments before transcription
  • Simulstreaming backend is now the default transcription backend
  • Added model recycling to optimize resource usage by removing whisper hooks at end of transcription
  • Diarization (diart backend) now correctly handles pauses and silences for better speaker turn detection
  • Default model changed to 'base'

Breaking Changes

  • VAC is now enabled by default (previously disabled) - may affect existing setups relying on VAC being off
  • Default transcription backend changed to simulstreaming (previously different)
  • Default model changed to 'base'

New Features

  • Voice Activity Control (VAC) enabled by default with --no-vac flag to disable
  • Simulstreaming backend as default with improved timestamp accuracy for audio over 30 seconds
  • Backend model recycling for resource optimization
  • Preload multiple backend models using --preloaded_model_count argument
  • Diarization handles pauses and silences correctly
  • Aligned time handling between backend and frontend
  • WebSocket buffering disabled during silent periods

Full Release Notes

- __Voice Activity Control (VAC) by Default__: VAC is now enabled by default to improve transcription accuracy by filtering out non-speech segments before processing transcription & diarization. You can disable it with the `--no-vac` flag.


- __Simulstreaming Backend Enhancements__:

  - The `simulstreaming` backend is now the default transcription backend.
  - Improved timestamp accuracy for audio segments longer than 30 seconds.
  - Backends models are now recycled to optimize resource usage, by removing whisper hooks at the end of a transcription
  - Added the ability to preload multiple backend models using the `--preloaded_model_count` argument, when several users are espected

- __Diarization with Silences__: The `diart` diarization backend now correctly handles pauses and silences, improving speaker turn detection.

- __Time Handling__: Aligned time handling between the backend and the frontend for better synchronization.


- __WebSocket Communication__: Buffering is disabled during silent periods.

- __Default Model__: The default model is now `base`.