BUTSpeechFIT/DiariZen

What it solves

DiariZen is a toolkit for speaker diarization—the process of determining "who spoke when" in an audio recording. It aims to provide high-accuracy diarization across diverse datasets without requiring domain adaptation for specific environments.

How it works

The toolkit is driven by AudioZen and Pyannote 3.1. It leverages self-supervised learning (SSL) and structured pruning to create compact, accurate models. The system supports training, pruning, and inference via a pipeline that can be integrated with Hugging Face pre-trained models.

Who it’s for

Researchers and developers working on speech processing, audio analysis, and automated transcription services who need an efficient, generalizable speaker diarization system.

Highlights

• High Performance: Outperforms Pyannote v3.1 on multiple benchmarks including AMI-SDM, AliMeeting, and VoxConverse.
• Model Compression: Includes recipes for structured pruning to remove redundant parameters while maintaining accuracy.
• Ease of Use: Provides a simple Python API for inference and supports saving results in RTTM format.
• Broad Support: Recently updated to support multi-channel WavLM for improved diarization.