README.md

metadata

base_model: pyannote/segmentation-3.0
library_name: transformers.js
license: mit

https://huggingface.co/pyannote/segmentation-3.0 with ONNX weights to be compatible with Transformers.js.

Transformers.js (v3) usage

import { AutoProcessor, AutoModelForAudioFrameClassification, read_audio } from '@xenova/transformers';

// Load model and processor
const model_id = 'onnx-community/pyannote-segmentation-3.0';
const model = await AutoModelForAudioFrameClassification.from_pretrained(model_id);
const processor = await AutoProcessor.from_pretrained(model_id);

// Read and preprocess audio
const url = 'https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/mlk.wav';
const audio = await read_audio(url, processor.feature_extractor.config.sampling_rate);
const inputs = await processor(audio);

// Run model with inputs
const { logits } = await model(inputs);
// {
//   logits: Tensor {
//     dims: [ 1, 767, 7 ],  // [batch_size, num_frames, num_classes]
//     type: 'float32',
//     data: Float32Array(5369) [ ... ],
//     size: 5369
//   }
// }

const result = processor.post_process_speaker_diarization(logits, audio.length);
// [
//   [
//     { id: 0, start: 0, end: 1.0512535626298245, confidence: 0.8220156481664611 },
//     { id: 2, start: 1.0512535626298245, end: 2.3398869619825127, confidence: 0.9008811707860472 },
//     ...
//   ]
// ]

// Display result
console.table(result[0], ['start', 'end', 'id', 'confidence']);
// ┌─────────┬────────────────────┬────────────────────┬────┬─────────────────────┐
// │ (index) │ start              │ end                │ id │ confidence          │
// ├─────────┼────────────────────┼────────────────────┼────┼─────────────────────┤
// │ 0       │ 0                  │ 1.0512535626298245 │ 0  │ 0.8220156481664611  │
// │ 1       │ 1.0512535626298245 │ 2.3398869619825127 │ 2  │ 0.9008811707860472  │
// │ 2       │ 2.3398869619825127 │ 3.5946089560890773 │ 0  │ 0.7521651315796233  │
// │ 3       │ 3.5946089560890773 │ 4.578039708226655  │ 2  │ 0.8491978128022479  │
// │ 4       │ 4.578039708226655  │ 4.594995410849717  │ 0  │ 0.2935352600416393  │
// │ 5       │ 4.594995410849717  │ 6.121008646925269  │ 3  │ 0.6788051309866024  │
// │ 6       │ 6.121008646925269  │ 6.256654267909762  │ 0  │ 0.37125512393851134 │
// │ 7       │ 6.256654267909762  │ 8.630452635138397  │ 2  │ 0.7467035186353542  │
// │ 8       │ 8.630452635138397  │ 10.088643060721703 │ 0  │ 0.7689364814666032  │
// │ 9       │ 10.088643060721703 │ 12.58113134631177  │ 2  │ 0.9123324509131324  │
// │ 10      │ 12.58113134631177  │ 13.005023911888312 │ 0  │ 0.4828358177572041  │
// └─────────┴────────────────────┴────────────────────┴────┴─────────────────────┘

Torch → ONNX conversion code:

# pip install torch onnx https://github.com/pyannote/pyannote-audio/archive/refs/heads/develop.zip
import torch
from pyannote.audio import Model

model = Model.from_pretrained(
  "pyannote/segmentation-3.0", 
  use_auth_token="hf_...", # <-- Set your HF token here
).eval()

dummy_input = torch.zeros(2, 1, 160000)
torch.onnx.export(
    model,
    dummy_input,
    'model.onnx',
    do_constant_folding=True,
    input_names=["input_values"],
    output_names=["logits"],
    dynamic_axes={
        "input_values": {0: "batch_size", 1: "num_channels", 2: "num_samples"},
        "logits": {0: "batch_size", 1: "num_frames"},
    },
)

---

Note: Having a separate repo for ONNX weights is intended to be a temporary solution until WebML gains more traction. If you would like to make your models web-ready, we recommend converting to ONNX using 🤗 Optimum and structuring your repo like this one (with ONNX weights located in a subfolder named onnx).