0.2.9

QuentinFuxa/WhisperLiveKit0.2.9Sep 11, 2025by QuentinFuxa

AI Summary

Release 0.2.9 introduces a beta live translation feature using NLLB, adds raw PCM input support and a punctuation split toggle, while fixing several GPU/CUDA issues with simulstreaming and improving overall stability.

Key Highlights

Beta live translation using NLLB with --target-language argument (may slow CPU transcription)
Fixed faster whisper encoder GPU/CUDA issues for simulstreaming
Added raw PCM input support via --pcm-input option
Fixed exponentially growing simulstreaming silence timer
Added --disable-punctuation-split for languages like Korean

New Features

Live translation using NLLB (beta) with --target-language argument
--disable-punctuation-split argument for non-space-separated languages
--pcm-input option for raw PCM audio input

Full Release Notes

## New Features
- **[BETA]** Live translation using NLLB - use `--target-language` argument. /!\ Can slow down transcription speed on CPU only devices. Improvements to come.
<div align="center">
<img width="600" src="https://github.com/user-attachments/assets/ed7ba5c6-909a-4109-8bf0-20f70686c595" />
</div>

- Added `--disable-punctuation-split` argument (useful for languages like Korean)
- Added support for raw PCM input using `--pcm-input` option

## Bug Fixes
- Fixed issue preventing faster whisper encoder from running on GPUs with CUDA for simulstreaming
- Fixed exponentially growing simulstreaming silence timer
- Fixed CUDA issue when using diarization with sortformer backend
- Fixed scrolling UX with sticky header controls
- Fixed simulstreaming `--preload-model-count` argument in CLI
- Fixed FFmpeg read timeout that frequently appeared after tasks ended


**Thank you to all contributors and people who reported issues and provided feedback.**