v0.2.19

QuentinFuxa/WhisperLiveKitv0.2.19Feb 23, 2026by QuentinFuxa

AI Summary

This release introduces the Voxtral backend with two variants: voxtral-mlx for native Apple Silicon performance (0.18-0.32x real-time) and voxtral (HF) for HuggingFace transformers on Linux/GPU. It also adds a comprehensive offline benchmark harness for testing all backends with WER, RTF, and timestamp accuracy metrics.

Key Highlights

Voxtral-mlx native MLX backend for Apple Silicon with 100+ language support
Voxtral (HF) HuggingFace transformers backend for Linux/GPU
New offline benchmark harness with JSON export capability
Full benchmark report with tables, charts, and recommendations
Fixed silence double-counting in audio processor
Fixed RTF inflation in metrics collector

New Features

Voxtral MLX backend (Apple Silicon)
Voxtral HuggingFace backend (Linux/GPU)
Offline benchmark harness (test_backend_offline.py --benchmark)
Benchmark report (BENCHMARK.md)

Full Release Notes

## Voxtral backend & benchmarks

### New: Voxtral backend

- **voxtral-mlx**: Native MLX backend for Apple Silicon. Runs at 0.18-0.32x real-time, handles 100+ languages with automatic language detection. No extra dependencies needed on macOS.
- **voxtral (HF)**: HuggingFace transformers backend for Linux/GPU. Requires `pip install transformers torch`.

### Benchmarks

New offline benchmark harness (`test_backend_offline.py --benchmark`) that runs all installed backends and computes WER, RTF, and timestamp accuracy against ground truth transcripts. Results exportable as JSON.

Full benchmark report in [BENCHMARK.md](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/BENCHMARK.md) with tables, charts, and recommendations for every backend/policy/model combination.

### Bug fixes

- Fixed silence double-counting in the audio processor
- Fixed median calculation for even-length lists in timestamp accuracy
- Fixed RTF inflation in metrics collector (was using wall-clock time instead of ASR processing time)