What's Changed
Major Update
- New speaker diarization feature via diar_sortformer_4spk-v1
- You can download my ONNX export here
Sortformer (end-to-end neural network):
- Audio → NN → 4 Speaker Probabilities (0.0-1.0) → Just threshold > 0.5 → Model outputs (fixed 4 channels)
- The model runs very fast on the CPU (tested on my Mac M3 so far), but additional processing may be required for very long audio files to avoid RAM bottlenecks.
See example: (diarization + tdt)
examples/diarization.rs
Note, Sortformer currently features flag only:
parakeet-rs = { version = "0.2.1", features = ["sortformer"] }
PRs:
- Feat add diar sortformer 4spk v1 for speaker diary by @altunenes in #17
Full Changelog: v0.1.8...v0.2.1