polyvoice 0.7.0

Speaker diarization for Python — who spoke when. Rust + ONNX, no Python runtime overhead. K-means/AHC clustering, overlap detection.

polyvoice

Speaker diarization for Python — who spoke when.

Rust-powered, ONNX-based speaker diarization that runs on CPU, fits in 30 MB, and requires zero Python runtime overhead. Pipeline v2 with K-means auto-k clustering and overlap detection.

Install

pip install polyvoice

Requires Python 3.9+.

Quick start

import polyvoice

# Models auto-download on first run (~30 MB)
pipeline = polyvoice.Pipeline.balanced()

result = pipeline.run(samples, sample_rate=16000)

print(f"Speakers: {result['num_speakers']}")
for turn in result["turns"]:
    print(f"Speaker {turn['speaker']}: {turn['start']:.1f}s - {turn['end']:.1f}s")

API

• polyvoice.Pipeline.balanced(models_cache=None) — balanced accuracy / speed.
• polyvoice.Pipeline.mobile(models_cache=None) — smaller, faster model.
• pipeline.run(samples, sample_rate) → dict with num_speakers and turns.

Performance

Pipeline	VoxConverse DER	Model size
Hybrid + K-means	14.12%	~30 MB

See the full repository for Rust / C / CLI APIs, benchmarks, and development docs.