QuentinFuxa/WhisperLiveKit 0.2.14

0.2.14.post1

on GitHub

New Features

• When using NLLW translation, the buffer is now displayed in light grey in the frontend.
• CLI / Parameters:

  • Can now be called with wlk instead of / in addition to whisperlivekit-server.

  • --task translate flag renamed to --direct-english-translation.

  • Backend parameters have been rethought:

    • --backend-policy sets the streaming strategy:
      simulstreaming uses AlignAtt SimulStreaming,
      localagreement uses the LocalAgreement policy.
    • --backend is now the Whisper implementation selector (SimulStreaming: encoder only; LocalAgreement: full encoder+decoder). auto selects the best config based on the system.
    • --model-path now accepts direct Hugging Face repo names (e.g. qanastek/whisper-tiny-french-cased, KBLab/kb-whisper-tiny, etc.).

  • For custom models, scripts/determine_alignment_heads.py determines the best alignment heads for SimulStreaming.

Bug Fixes

• Fixed flickering between speakers when using diarization.
• Correct task resolution at the end of the stream.
• Fixed custom model handling for some models (e.g. #269).
• 0.2.14.post1: Improve silence handling, by producing remaining non-validated tokens when silence starts