Pyannote ONNX Extended

A pure ONNX Runtime implementation of the Pyannote Speaker Diarization 3.1 (multi-speaker) pipeline.

This project removes the heavy PyTorch dependency for inference, making it lightweight, fast, and easy to deploy.

Based on the pyannote-audio models and inspired by pyannote-onnx.

Key Features

• Pure ONNX Runtime: No PyTorch required for inference.
• Robust Overlap Handling: Implements "Average Stitching" to handle overlapping speech segments smoothly across sliding windows.
• Two-Stage Clustering: Uses a specialized clustering approach where stable "long" segments defined the speakers, and "short" segments are assigned to the nearest speaker. This significantly improves stability for short utterances.
• Lightweight: Minimal dependencies compared to the full PyTorch pipeline.

Exporting Models (Optional)

If you'd like to export the PyTorch models to ONNX format by yourself, you can do so by running the following command:

pip install -r requirements.txt

You will need a Hugging Face token with access to pyannote/speaker-diarization-3.1.

python export_onnx.py --use_auth_token YOUR_HF_TOKEN

This will create a models_onnx folder containing:

• segmentation.onnx
• embedding.onnx

Installation

pip install .

Usage

from onnx_pyannote import ONNXSpeakerDiarization

# Initialize the pipeline
pipeline = ONNXSpeakerDiarization(
    model_name="speaker-diarization-3.1",
    providers=['CUDAExecutionProvider', 'CPUExecutionProvider'] # Use CUDA if available
)

# Process an audio file
audio_path = "path/to/your/audio.wav"
annotation = pipeline(audio_path)

# Print result
for turn, _, speaker in annotation.itertracks(yield_label=True):
    print(f"start={turn.start:.1f}s stop={turn.end:.1f}s speaker={speaker}")