polyvoice

Speaker diarization for Rust — who spoke when, on CPU, without Python.

Beta-quality, ONNX-powered, ~30 MB. Embeds into any Rust app, with Python, C, and CLI bindings.

Speaker_0: 0.0s - 12.3s
Speaker_1: 14.1s - 28.7s
Speaker_0: 31.2s - 45.0s

Like-for-like (collar 0, overlap-scored) VoxConverse-test DER is 15.4% (v2+VBx default) vs pyannote 3.1's 11.3% — a few DER points traded for a CPU-only, MIT, ungated engine that needs no Python — see Benchmarks.

Install

Pre-built CLI binary

Download the latest binary for your platform from the GitHub Releases page:

# Linux x86_64
curl -L -o polyvoice https://github.com/ekhodzitsky/polyvoice/releases/latest/download/polyvoice-linux-x86_64
chmod +x polyvoice
sudo mv polyvoice /usr/local/bin/

# Linux ARM64
curl -L -o polyvoice https://github.com/ekhodzitsky/polyvoice/releases/latest/download/polyvoice-linux-aarch64
chmod +x polyvoice
sudo mv polyvoice /usr/local/bin/

# macOS (Apple Silicon)
curl -L -o polyvoice https://github.com/ekhodzitsky/polyvoice/releases/latest/download/polyvoice-macos-arm64
chmod +x polyvoice
sudo mv polyvoice /usr/local/bin/

Docker

docker run --rm -v "$(pwd):/work" ghcr.io/ekhodzitsky/polyvoice:latest diarize /work/meeting.wav --output /work/meeting.rttm

Rust library

cargo add polyvoice --features "onnx,download"

Python

pip install polyvoice

From source

# CLI (WAV 16 kHz mono input). Feature `cli` includes VBx (default clusterer).
cargo install polyvoice --features cli

# CLI + any-format audio (mp3/flac/ogg/m4a/aac, any sample rate → 16 kHz mono)
cargo install polyvoice --features "cli,audio-io"

Usage

use polyvoice::models::ModelRegistry;
use polyvoice::pipeline_v2::{ClustererKind, Pipeline, PipelineConfig};
use polyvoice::types::{Profile, SampleRate};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut cfg = PipelineConfig {
        profile: Profile::Balanced,
        clusterer: ClustererKind::Vbx, // or Ahc { threshold: 0.45 }
        ..PipelineConfig::default()
    };
    // VBx needs PLDA weights: cfg.vbx_plda_dir = Some("data/vbx-plda".into());
    // or POLYVOICE_VBX_PLDA_DIR in the environment.
    let pipeline = Pipeline::builder()
        .config(cfg)
        .with_models_from(ModelRegistry::default()?) // models auto-download on first run
        .build()?;

    // Pipeline-ready mono 16 kHz (with feature `audio-io`, also mp3/flac/… + resample).
    // Without that feature, load_audio accepts 16 kHz WAV only.
    let (samples, sr) = polyvoice::wav::load_audio(std::path::Path::new("meeting.wav"))?;
    let result = pipeline.run(&samples, SampleRate::new(sr).ok_or("bad sample rate")?)?;

    for turn in &result.turns {
        println!("{}: {:.1}s - {:.1}s", turn.speaker, turn.time.start, turn.time.end);
    }
    Ok(())
}

polyvoice download-models --profile balanced
# Default path: pipeline v2 + VBx (needs PLDA)
export POLYVOICE_VBX_PLDA_DIR=/path/to/vbx-plda   # see docs/vbx-plda-release.md
polyvoice diarize meeting.wav --output meeting.rttm
# Without PLDA: --clusterer ahc   |   old path: --legacy
# With a build that includes `audio-io`:
# polyvoice diarize meeting.mp3 --output meeting.rttm

Python usage and the full API live on docs.rs.

Why polyvoice

• Maintained, pure-Rust, streaming-capable. The popular sherpa-rs bindings are archived; polyvoice is an actively-maintained, pure-Rust diarization path (ONNX via ort, no C++ toolkit) with first-class streaming.
• One library, four surfaces. Rust + Python + C FFI + CLI from a single crate.
• CPU-first, ~30 MB, MIT. No GPU, no Python runtime, no gated model access.

It is not the accuracy leader — like-for-like (collar 0, overlap-scored) VoxConverse-test DER is 15.4% (v2+VBx default) versus 11.3% for pyannote 3.1. It trades those DER points for deployability: a pure-Rust, CPU, MIT, ungated engine (pyannote's weights are gated behind an HF token) with four bindings and streaming.

How it works

audio (f32 PCM)
  → VAD / Powerset segmentation
  → WeSpeaker embeddings
  → clustering (AHC / K-means / NME-SC, automatic speaker count)
  → speaker turns

Streaming (OnlineDiarizer) and batch (OfflineDiarizer), with a single-speaker guard so quiet or single-voice audio does not hallucinate clusters.

Documentation

• Benchmarks — collar-disclosed DER numbers and provenance
• Production readiness — deployment guidance (GO / NO-GO)
• Migrating from 0.5 · Glossary
• Contributing · Changelog

License

MIT

Name: this project is polyvoice — speaker diarization for Rust, unrelated to ByteDance's "PolyVoice" speech-translation research.