Starstruk Logo MarkStarstruk

0.2.10

QuentinFuxa/WhisperLiveKit0.2.10Sep 17, 2025by QuentinFuxa

AI Summary

This release focuses on enhancing the Live Translation with NLLB feature by adding backend flexibility (switchable via --nllb-backend) and configurable model sizes (600M to 1.3B via --nllb-size). It also introduces frontend AudioWorklet support for PCM input and includes several bug fixes related to translation performance, websocket timeouts on low-end devices, and buffer positioning issues.

Key Highlights

  • NLLB backend flexibility - switch between backends using --nllb-backend for improved performance
  • Configurable NLLB model size - adjust from 600M to 1.3B using --nllb-size parameter
  • Frontend AudioWorklet support - automatically used when --pcm-input is enabled
  • Translation now runs in dedicated asyncio thread to prevent transcription/diarization slowdowns
  • Resolved websocket timeout issues on low-end devices

New Features

  • NLLB backend switching via --nllb-backend parameter for backend flexibility
  • NLLB model size configuration via --nllb-size parameter (600M to 1.3B)
  • Frontend AudioWorklet support for PCM input environments without ffmpeg

Full Release Notes

## New Features

- Live Translation with NLLB:
    - Backend flexibility: Add possibility to switch to the 'transformers' backend using `--nllb-backend` for improved performance on certain systems. See [here](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/available_models.md) for guidance on choosing the optimal backend.
    - Configurable model size: Adjust NLLB model size from 600M to 1.3B using `--nllb-size` parameter for enhanced translation accuracy.

- Frontend AudioWorklet support: The frontend now automatically uses AudioWorklet instead of MediaRecorder when `--pcm-input` is enabled, useful for environments without ffmpeg or for custom implementations.

## Bug Fixes

-  Fixed `--warmup-file` parameter compatibility with backends like faster-whisper
- Translation now runs in a dedicated asyncio thread to prevent transcription and diarization slowdowns
- Translation processing now correctly manages silent periods
- Resolved websocket timeout issues on low-end devices
- Corrected transcription/translation buffer positioning that was appearing below the translation output

**Thank you to all contributors and community members who reported issues and provided valuable feedback.**