Starstruk Logo MarkStarstruk

0.2.8

QuentinFuxa/WhisperLiveKit0.2.8Sep 2, 2025by QuentinFuxa

AI Summary

This release focuses on performance improvements for the simulstreaming backend by defaulting to MLX-Whisper or Faster-Whisper encoders with AlignAtt policy, while also reducing vRAM usage through shared encoder loading. The release also adds Python 3.14/3.15 compatibility, removes the Triton <3 requirement, and introduces frontend enhancements including a microphone picker and simplified deployment via a single inline HTML file.

Key Highlights

  • Removed Triton <3 requirement and added Python 3.14/3.15 compatibility
  • Simulstreaming backend now defaults to MLX-Whisper/Faster-Whisper encoders with AlignAtt policy for increased speed
  • Reduced vRAM usage - encoders loaded once and shared, only Whisper decoder loaded when using different encoder
  • Added microphone picker to frontend
  • Fixed warmup errors and language auto-detection issues including 'auto' language with 'translation' task

New Features

  • Microphone picker for selecting input device
  • Single inline HTML file for UI (simplified deployment)
  • Auto-detection of language for warmup if not specified
  • --disable-fast-encoder flag to disable the new fast encoder defaults
  • Added pip timeout and retries in Dockerfile for Torch/TorchVision/TorchAudio installation

Full Release Notes

### Dependency and Compatibility Changes

- Removed Triton <3 requirement
- Tested compatibility with Python 3.14 and 3.15

### Performance Improvements

- Simulstreaming backend now defaults to MLX-Whisper (if available) or Faster-Whisper (if available) encoders, paired with Whisper cross-attention and decoder using an AlignAtt policy, for increased speed. Can be disabled using `--disable-fast-encoder`
- Encoders are loaded once and shared in Simulstreaming, reducing vRAM usage
- Only the decoder of Whisper is loaded when using a different encoder, reducing vRAM usage

### Frontend Enhancements

- Added a microphone picker
- Loads the UI as a single inline HTML file (instead of separate CSS, JS, SVGs and HTML files) for simplified deployment

### Bug Fixes and Improvements

- Resolved warmup error when no connection is provided or when the language is set to auto
- Added pip timeout and retries in Dockerfile when installing Torch/TorchVision/TorchAudio
- Fixed issue where an exception is raised when language is set to 'auto' and task is set to 'translation'
- Enabled auto-detection of language for warmup if not specified