Diart

Diart (juanmc2005/diart) is the official implementation of the Interspeech 2021 paper 'Overlap-aware low-latency online speaker diarization based on end-to-end local segmentation' (Coria, Bredin, Ghannay, Rosset). It's the go-to toolkit when diarization has to happen on live audio rather than files: a rolling 5-second buffer updated every 500ms, combining incremental clustering with local end-to-end segmentation that is explicitly overlap-aware — it can report two people talking at once, not just assign the segment to one of them. The clever bits are a statistics-pooling layer that down-weights frames where the segmentation predicts simultaneous speakers, and cannot-link constraints that stop the clustering step from wrongfully merging two local speakers. Latency is a dial you turn: 500ms to 5s, with a systematic latency-vs-performance analysis in the paper (evaluated on AMI, DIHARD and VoxConverse). Under the hood it wires together pyannote segmentation and embedding models (Hugging Face-gated) with a CLI (diart.stream /path/to/audio.wav or diart.stream microphone), a Python API, WebSocket audio sources, and RTTM output. Licensed MIT; the models it pulls are pyannote's, so their HF gating applies. If you need who-spoke-when on a live meeting feed rather than a finished recording, this is the classic answer.

• Paper evaluates latency/accuracy trade-off on AMI, DIHARD and VoxConverse (author-reported, Interspeech 2021)\n- No independent published DER benchmark or head-to-head comparison found