---
license: mit
language:
  - en
  - fr
  - multilingual
library_name: diarization-js
tags:
  - speaker-diarization
  - diarization
  - audio
  - speech
  - pyannote
  - onnx
  - onnxruntime
  - webgpu
pipeline_tag: voice-activity-detection
---

# diarization-js · community-1 artifacts

ONNX-exported model weights and PLDA parameters bundled for the
[`diarization-js`](https://www.npmjs.com/package/diarization-js) JavaScript
library — a port of `pyannote/speaker-diarization-community-1` that runs
fully in the browser (WebGPU / WASM) or in Node, no Python required.

> ⚠️ This repository hosts **artifacts only** (model files). The pipeline
> code lives in the npm package `diarization-js`.

# Live demo
['test a full scale demo in browser'](https://diarization.asimov.works/)

## Contents

| File | Size | Purpose | Origin | License |
|------|------|---------|--------|---------|
| `segmentation-3.0.onnx`   | ~6 MB  | Powerset multi-label segmentation backbone | [`pyannote/segmentation-3.0`](https://huggingface.co/pyannote/segmentation-3.0) | MIT |
| `embedding-resnet34.onnx` | ~26 MB | ResNet34 speaker embedding extractor | [`pyannote/wespeaker-voxceleb-resnet34-LM`](https://huggingface.co/pyannote/wespeaker-voxceleb-resnet34-LM) | CC-BY-4.0 |
| `plda-params-vbx.json`    | ~1 MB  | PLDA parameters for VBx clustering | [`pyannote/pyannote-audio`](https://github.com/pyannote/pyannote-audio) | MIT |

These three files are functionally equivalent to the upstream
`pyannote/speaker-diarization-community-1` pipeline at the time of export.
Validated end-to-end against the Python reference (DER = 1.73% on
VoxConverse v0.3 dev).

## Usage

Install the runtime:

```bash
npm install diarization-js onnxruntime-web   # browser
npm install diarization-js onnxruntime-node  # Node
```

Load and run:

```ts
import * as ort from "onnxruntime-web/webgpu";
import { DiarizationPipeline, ensureArtifacts } from "diarization-js";

const artifacts = await ensureArtifacts(); // fetches this repo, caches
const pipeline = await DiarizationPipeline.load({
  ort,
  ...artifacts,
});

const result = await pipeline.run(float32Audio16k);
console.log(result.segments);
```

In Node the artifacts are cached on disk under `~/.cache/diarization-js/`.
In the browser the HTTP cache (max-age headers from huggingface.co) handles
deduplication — first load fetches ~33 MB, subsequent loads are local.

## Self-hosting

The library never hard-requires huggingface.co. Pass any base URL that
serves the same three filenames to `ensureArtifacts({ source: "..." })`,
or load the bytes yourself and pass them directly to
`DiarizationPipeline.load`.

## Citing

If you use this package in academic work, please cite the upstream
pyannote.audio papers:

```bibtex
@inproceedings{Bredin2023,
  title={pyannote.audio 2.1 speaker diarization pipeline},
  author={Bredin, Hervé},
  booktitle={Interspeech},
  year={2023}
}
```

## License

- **Code** (diarization-js npm package): MIT
- **Weights** in this repository: see per-file table above. Redistributed
  here under the upstream licenses; original copyright holders retain
  their rights.