speakrs

Speaker diarization in Rust. Runs 312-912x realtime on Apple Silicon (20-50x faster than pyannote) and 50-121x on CUDA (2-7x faster than pyannote), matching pyannote accuracy.

speakrs implements the full pyannote community-1 pipeline in Rust: segmentation, powerset decode, aggregation, binarization, embedding, PLDA, and VBx clustering, plus temporal smoothing during reconstruction. There is no Python dependency. Inference runs on ONNX Runtime or native CoreML, and all post-processing stays in Rust.

On VoxConverse dev (216 files), speakrs CoreML achieves 7.1% DER at 529x realtime vs pyannote's 7.2% at 24x. On the test set (232 files) speakrs matches pyannote at 11.1% DER while running 27x faster. On CUDA, speakrs matches or beats pyannote DER on all datasets at 2-7x the speed. See benchmarks/ for full results across 8 datasets.

Overview

Speaker diarization in Rust

speakrs implements the full pyannote community-1 speaker diarization pipeline: segmentation, powerset decode, aggregation, binarization, embedding, PLDA, and VBx clustering. There is no Python dependency. Inference runs on ONNX Runtime or native CoreML, and all post-processing stays in Rust.

Usage

# Apple Silicon (CoreML)
speakrs = { version = "0.2", features = ["coreml"] }

# NVIDIA GPU
speakrs = { version = "0.2", features = ["cuda"] }

# CPU only (default)
speakrs = "0.2"

Quick start

use speakrs::{ExecutionMode, OwnedDiarizationPipeline};

fn main() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
    let mut pipeline = OwnedDiarizationPipeline::from_pretrained(ExecutionMode::CoreMl)?;

    let audio: Vec<f32> = load_your_mono_16khz_audio_here();
    let result = pipeline.run(&audio)?;

    print!("{}", result.rttm("my-audio"));
    Ok(())
}

Speaker turns

use speakrs::pipeline::{FRAME_STEP_SECONDS, FRAME_DURATION_SECONDS};

let result = pipeline.run(&audio)?;

for segment in result.discrete_diarization.to_segments(FRAME_STEP_SECONDS, FRAME_DURATION_SECONDS) {
    println!("{:.3} - {:.3}  {}", segment.start, segment.end, segment.speaker);
}

Background queue

For processing many files, [QueueSender] and [QueueReceiver] run a background worker that auto-batches requests for cross-file optimizations. Push files from any thread, receive results as they complete:

use speakrs::{ExecutionMode, OwnedDiarizationPipeline, QueuedDiarizationRequest};

let pipeline = OwnedDiarizationPipeline::from_pretrained(ExecutionMode::CoreMl)?;
let (tx, rx) = pipeline.into_queued()?;

// producer: push files as they arrive from another thread
std::thread::spawn(move || {
    for (file_id, audio) in receive_files() {
        tx.push(QueuedDiarizationRequest::new(file_id, audio)).unwrap();
    }
    // tx drops when the producer is done, ending the rx loop
});

// results arrive while files are still being pushed,
// the loop exits once tx is dropped and all queued work is done
for result in rx {
    let result = result?;
    print!("{}", result.result?.rttm(&result.file_id));
}

Local models

For offline or airgapped usage, load models from a local directory:

use std::path::Path;
use speakrs::{ExecutionMode, OwnedDiarizationPipeline};

let mut pipeline = OwnedDiarizationPipeline::from_dir(
    Path::new("/path/to/models"),
    ExecutionMode::Cpu,
)?;
let result = pipeline.run(&audio)?;

Pipeline

Audio (16kHz f32)
  │
  ├─ Segmentation ──────→ raw 7-class logits per 10s window
  │   (ONNX or CoreML)
  │
  ├─ Powerset Decode ───→ 3-speaker soft/hard activations
  │
  ├─ Overlap-Add ───────→ Hamming-windowed aggregation across windows
  │
  ├─ Binarize ──────────→ hysteresis thresholding + min-duration filtering
  │
  ├─ Embedding ─────────→ 256-dim WeSpeaker vectors per segment
  │   (ONNX or CoreML)
  │
  ├─ PLDA Transform ────→ 128-dim whitened features
  │
  ├─ VBx Clustering ────→ Bayesian HMM speaker assignments
  │
  ├─ Reconstruct ───────→ map clusters back to frame-level activations (temporal smoothing)
  │
  └─ Segments ──────────→ RTTM output

macOS / iOS (CoreML)

Requires the coreml Cargo feature. Uses Apple's CoreML framework for GPU/ANE-accelerated inference.

Execution modes

coreml-fast uses a wider step (2s instead of 1s) to get about 1.5x more speed. That follows the same throughput-first tradeoff SpeakerKit uses on Apple hardware. It matches coreml on most clips, but on some inputs the coarser step loses temporal resolution at speaker boundaries.

Benchmarks

All benchmarks on Apple M4 Pro, macOS 26.3, evaluated on VoxConverse dev (216 files, 1217.8 min, collar=0ms):

On VoxConverse test (232 files, 2612.2 min), coreml matches pyannote at 11.1% DER while running at 631x realtime vs pyannote's 23x.

CoreML results may differ slightly from ONNX CPU, the two runtimes apply different graph optimizations (operator fusion, reduction order) that change floating-point rounding, even on CPU in FP32. Apple's own measurements show ~96 dB SNR between CoreML FP32 and source frameworks (typed execution docs). See benchmarks/ for full results across multiple datasets.

Choosing a mode

The accuracy gap between coreml and coreml-fast depends on the type of audio. On meeting recordings with orderly turn-taking (AMI), CoreML Fast matches CoreML within 0.4% DER. On broadcast content with frequent speaker changes (VoxConverse), the gap is ~0.3%. On some datasets like Earnings-21, CoreML Fast actually beats CoreML on DER.

coreml-fast never misses speech or hallucinates extra speech. The only extra errors are misattributing speech to the wrong speaker near turn boundaries, because the 2s step gives fewer data points to pinpoint where one speaker stops and another starts.

See benchmarks/ for full results across all datasets.

CPU & CUDA (Linux, Windows, macOS)

Works on any platform with ONNX Runtime. No special Cargo features needed for CPU. Enable the cuda feature for NVIDIA GPU acceleration.

Execution modes

Benchmarks

NVIDIA RTX 4090, AMD EPYC 7B13, evaluated on VoxConverse dev (216 files, 1217.8 min, collar=0ms):

On VoxConverse test (232 files, 2612.2 min, L40S), cuda matches pyannote at 11.1% DER at 50x realtime vs pyannote's 18x.

Models

Models download automatically on first use from avencera/speakrs-models on HuggingFace. If you want a custom model directory, set SPEAKRS_MODELS_DIR.

Public API

Why not pyannote-rs?

pyannote-rs is another Rust diarization crate, but it uses a simpler pipeline instead of the full pyannote algorithm:

On the VoxConverse dev set, using the 33 files where pyannote-rs produces output (186 min, collar=0ms):

pyannote-rs produces 0 segments on 183 out of 216 VoxConverse files. Its segments only close on speech to silence transitions, so continuous speech without silence gaps yields no output. The 33 files above are the subset where it produces at least 5 segments.

Note: pyannote-rs's README says it uses wespeaker-voxceleb-resnet34-LM, but their build instructions and GitHub release only ship wespeaker_en_voxceleb_CAM++.onnx. There is no ONNX export of ResNet34-LM. The HuggingFace repo only contains pytorch_model.bin. The benchmark here uses pyannote-rs exactly as documented in their setup instructions.

Features

• online (default) — automatic model download from HuggingFace via [ModelManager]
• coreml — native CoreML backend for Apple Silicon GPU/ANE acceleration
• cuda — NVIDIA CUDA backend via ONNX Runtime
• load-dynamic — load the CUDA runtime library at startup instead of static linking

Build requirements

This crate links OpenBLAS statically (via ndarray-linalg), which requires a C compiler. On most systems this is already available. The ONNX Runtime dependency (ort 2.0.0-rc.12) is pre-release.

Contributing

See CONTRIBUTING.md for local setup, model downloads, fixture generation, and the standard check commands used in this repo.

References

• pyannote-audio - Python reference implementation
• pyannote community-1 - VBx + PLDA pipeline
• SpeakerKit - Swift reference (same VBx architecture)