What's Changed
- Per-model streaming flush intervals: Whisper models keep a 10s flush window for optimal transcription quality (benchmarked: 13.4% WER at 5s vs 6.3% at 10s on large-v3-turbo), while Parakeet and Qwen models get a 5s flush for faster feedback (#195)
- Homebrew cask: Bumped to v1.32.0
Benchmark Tool
The benchmark tool now supports --chunked mode to compare flush intervals side by side, making it easier to evaluate streaming transcription quality across models and window sizes.
Full Changelog: v1.32.0...v1.32.1