# pyannote-speaker-diarization-3.1-mlx

First MLX port of pyannote-speaker-diarization-3.1 with byte-parity to the PyTorch reference. 2.5x faster than pyannote-MPS on Apple Silicon native.

## Install

```bash
uv add "pyannote-speaker-diarization-3.1-mlx @ git+https://gitea.tavportal.com/olivier/pyannote-speaker-diarization-3.1-mlx.git"
```

## Quickstart

```python
from pyannote_diarization_3_1_mlx import MlxDiarizationPipeline

pipeline = MlxDiarizationPipeline.from_pretrained("pyannote/speaker-diarization-3.1")
diarization = pipeline("audio.wav")

for turn, _, speaker in diarization.itertracks(yield_label=True):
    print(f"{turn.start:.1f}s - {turn.end:.1f}s {speaker}")
```

## Parity

| Evidence | MLX | Reference | Result |
| --- | --- | --- | --- |
| Cosine distance (200 cross-window pairs) | mean=0.763718 | pyannote-PyTorch mean=0.763718 | identical at 6 decimals |
| 5h10 bench | 173s / 44 speakers / 1.27 GB | pyannote-MPS 431s / 43 speakers / 1.72 GB | Cross-DER 0.076 |

## Architecture

SincNet → BiLSTM → Powerset(3,2) head + WeSpeaker ResNet34 speaker embedding + AgglomerativeClustering wrapper.

## Module Naming

The repository name is `pyannote-speaker-diarization-3.1-mlx`; the Python import is `pyannote_diarization_3_1_mlx`. The import name follows PEP 8 and embeds the pyannote model version so future 4.0 ports can co-install.

## Citation

This project ports the pyannote speaker diarization 3.1 pipeline architecture to MLX. Please cite the original pyannote.audio work when using this package:

```bibtex
@inproceedings{Plaquet23,
  author = {Alexis Plaquet and Hervé Bredin},
  title = {{Powerset multi-class cross entropy loss for neural speaker diarization}},
  booktitle = {Proc. INTERSPEECH 2023},
  year = {2023},
}
```

## Provenance

Extracted from MLX_CONVERTOR/src/mlxconv/diar at commit 5f9eafa. Maintained at https://gitea.tavportal.com/olivier/pyannote-speaker-diarization-3.1-mlx.

## License

MIT