speakrs models

Model files for speakrs, a Rust speaker diarization library that runs the pyannote community-1 pipeline natively on CPU, CUDA (NVIDIA) and Core ML (macOS), with ONNX Runtime for AMD GPUs. With the default online feature, speakrs downloads the files it needs on first use from a pinned revision of this repository.

These are exports and conversions of pyannote's models, not new or retrained models. The Core ML variants include FP16 and 8-bit-weight (W8A16) segmentation models.

Speed

VoxConverse dev (20.3 h), collar 0:

Hardware	Implementation	DER	Speed (x real time)
RTX 4090	speakrs `cuda`	7.0 %	1912x
RTX 4090	speakrs `cuda-fast`	7.4 %	3646x
RTX 5090	speakrs `cuda`	7.0 %	2922x
RTX 4090	pyannote community-1 (PyTorch)	7.2 %	32x
Apple M4 Pro	speakrs `coreml`	7.1 %	529x

Hardware details and the timing method are in the speakrs benchmarks.

Usage

speakrs = { version = "0.7", features = ["coreml"] } # macOS
speakrs = { version = "0.7", features = ["cuda"] }   # NVIDIA GPU
speakrs = { version = "0.7", features = ["cpu"] }    # CPU

use speakrs::{ExecutionMode, OwnedDiarizationPipeline};

let mut pipeline = OwnedDiarizationPipeline::from_pretrained(ExecutionMode::Cuda)?;
let result = pipeline.run(&audio)?; // mono 16 kHz f32 samples
print!("{}", result.rttm("my-audio"));

Python:

pip install speakrs

import speakrs

pipeline = speakrs.Pipeline.from_pretrained("cpu")  # or "coreml", "cuda"
print(pipeline.diarize_file("meeting.wav").rttm())

For offline setups, download this repository and set SPEAKRS_MODELS_DIR to its local path.

Files	Used by	Source
`segmentation-3.0.safetensors`	native CPU and CUDA	pyannote/segmentation-3.0
`wespeaker-multimask-tail.safetensors`	native CPU and CUDA	pyannote/wespeaker-voxceleb-resnet34-LM
`segmentation-3.0.mlmodelc`, `wespeaker.mlmodelc`	Core ML	the same two models, converted
`*.onnx`	ONNX Runtime (MIGraphX)	the same two models, exported
`plda_*.npy`	all modes, clustering	PLDA and VBx parameters from pyannote/speaker-diarization-community-1

License

Derived from pyannote models:

pyannote/segmentation-3.0: MIT License, copyright CNRS
pyannote/wespeaker-voxceleb-resnet34-LM: CC BY 4.0
pyannote/speaker-diarization-community-1: CC BY 4.0

Changes: exported to safetensors and ONNX, converted to Core ML (including FP16 and W8A16 variants), and PLDA parameters extracted for speakrs. Credit: pyannote.audio by Hervé Bredin and contributors, and WeSpeaker.