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.

Table of Contents

• Pipeline
• macOS / iOS (CoreML)
• CPU & CUDA (Linux, Windows, macOS)
• Usage
• Models
• Modules
• Why Not pyannote-rs?
• Contributing
• References

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

Mode	Backend	Step	Precision	Use case
coreml	Native CoreML	1s	FP32	Best accuracy (529x realtime)
coreml-fast	Native CoreML	2s	FP32	Best speed (912x realtime)

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):

Mode	DER	Time	RTFx
coreml	7.1%	138s	529x
coreml-fast	7.4%	169s	434x
pyannote community-1 (MPS)	7.2%	2999s	24x
SpeakerKit	7.8%	234s	312x

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

Mode	Backend	Step	Precision	Use case
cpu	ORT CPU	1s	FP32	Reference
cuda	ORT CUDA	1s	FP32	Best accuracy
cuda-fast	ORT CUDA	2s	FP32	Best speed

Benchmarks

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

Mode	DER	Time	RTFx
cuda	7.0%	1236s	59x
cuda-fast	7.4%	604s	121x
pyannote community-1 (CUDA)	7.2%	2312s	32x

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

Usage

Add to your Cargo.toml:

# Apple Silicon (CoreML)
speakrs = { version = "0.1", features = ["coreml"] }

# NVIDIA GPU
speakrs = { version = "0.1", features = ["cuda"] }

# CPU only (default)
speakrs = "0.1"

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(())
}
# fn load_your_mono_16khz_audio_here() -> Vec<f32> { unimplemented!() }

Speaker turns

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

let mut pipeline = OwnedDiarizationPipeline::from_pretrained(ExecutionMode::CoreMl)?;
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);
}
# Ok::<(), Box<dyn std::error::Error + Send + Sync>>(())

Background queue

For processing many files, QueuedDiarizationPipeline runs a background worker that auto-batches requests for cross-file optimizations:

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

let pipeline = OwnedDiarizationPipeline::from_pretrained(ExecutionMode::CoreMl)?;
let queue = pipeline.into_queued()?;

queue.push(QueuedDiarizationRequest::new("file1", audio1))?;
queue.push(QueuedDiarizationRequest::new("file2", audio2))?;

for result in queue {
    let diarization = result.result?;
    print!("{}", diarization.rttm(&result.file_id));
}
# Ok::<(), Box<dyn std::error::Error + Send + Sync>>(())

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)?;
# Ok::<(), Box<dyn std::error::Error + Send + Sync>>(())

CLI

# CoreML (best accuracy, requires coreml feature)
cargo run --release -p xtask --features coreml --bin diarize -- --mode coreml audio.wav

# CoreML Fast (fastest on Apple Silicon)
cargo run --release -p xtask --features coreml --bin diarize -- --mode coreml-fast audio.wav

# CPU
cargo run --release -p xtask --bin diarize -- --mode cpu audio.wav

# CUDA (NVIDIA GPU)
cargo run --release -p xtask --features cuda,load-dynamic --bin diarize -- --mode cuda audio.wav

See examples/README.md for more runnable examples, including airtime reporting and transcript speaker assignment.

Models

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

For development, just export-models exports the ONNX models and converts them to CoreML. That command requires Python through uv.

Public API

Module / Type	Description
OwnedDiarizationPipeline	Main entry point, owns models and runs diarization
QueuedDiarizationPipeline	Background worker with push/recv queue interface
DiarizationPipeline	Borrowed pipeline for manual model lifetime control
DiarizationResult	All outputs: segments, embeddings, clusters, RTTM
ExecutionMode	CPU, CoreML, CoreMLFast, CUDA, CUDAFast
PipelineConfig / RuntimeConfig	Tunable pipeline and hardware parameters
Segment	A single speaker turn with start/end times
ModelManager	Automatic model download from HuggingFace
inference	Segmentation and embedding model wrappers
segment	Segment conversion, merging, RTTM formatting

Why Not pyannote-rs?

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

	speakrs	pyannote-rs
Segmentation	Powerset decode → 3-speaker activations	Raw argmax on logits (binary speech/non-speech)
Aggregation	Hamming-windowed overlap-add	None (per-window only)
Binarization	Hysteresis + min-duration filtering	None
Embedding model	WeSpeaker ResNet34 (same as pyannote)	WeSpeaker CAM++ (only ONNX model they ship)
Clustering	PLDA + VBx (Bayesian HMM)	Cosine similarity with fixed 0.5 threshold
Speaker count	VBx EM estimation	Capped by max_speakers parameter
pyannote parity	Bit-exact on CPU/CUDA	No, different algorithm and different embedding model

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

	DER	Missed	False Alarm	Confusion
speakrs CoreML	11.5%	3.8%	3.6%	4.1%
pyannote-rs	80.2%	34.9%	7.4%	37.9%

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.

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)