0.2.7

QuentinFuxa/WhisperLiveKit0.2.7Aug 27, 2025by QuentinFuxa

AI Summary

Release 0.2.7 introduces major diarization improvements with Sortformer as the new default backend, delivering 6x faster processing and significantly better speaker detection while reducing memory usage through shared model loading.

Key Highlights

Sortformer is now the default diarization backend, replacing Diart
6x faster processing: reduced latency from ~2s to ~0.3s on CPU
Significantly improved speaker detection (max 4 speakers)
Shared model architecture reduces memory footprint across users/instances
Enhanced alignment between transcription and diarization results

Breaking Changes

Default diarization backend changed from Diart to Sortformer
Speaker detection now limited to maximum 4 speakers

New Features

SortformerDiarization as the new default diarization backend
SortformerDiarizationOnline for per-user speaker tracking (caches, frames)
Shared model loading to reduce memory footprint
Improved time and token synchronization between transcription and diarization

Full Release Notes

## 0.2.7: Diarization Improvements

- **New default backend**: Sortformer is now the default diarization backend, replacing Diart
- **6x faster processing**: Reduced latency from ~2s to ~0.3s on CPU
- **Significantly improved speaker detection** (**Constraint**: Currently supports up to 4 speakers maximum)
- **Shared model loading**: A single Sortformer model `SortformerDiarization` is now shared across users and instances to reduce memory footprint. Speaker caches, frames, etc. are handled per user in `SortformerDiarizationOnline`
- **Enhanced alignment**: Improved time and token synchronization between transcription and diarization results