ediar / README.md
multimodalart's picture
multimodalart HF Staff
Upload folder using huggingface_hub
8f92284 verified
|
Raw History Blame Contribute Delete
1.83 kB
metadata
title: eDIAR — Open-Speaker Streaming Diarization
emoji: 🎧
colorFrom: gray
colorTo: green
sdk: gradio
sdk_version: 6.32.0
app_file: app.py
python_version: '3.12'
short_description: Training-free streaming diarization for meetings via e-means
startup_duration_timeout: 1h

eDIAR: Training-Free Open-Speaker Streaming Diarization

Demo of eDIAR (paper) — training-free, open-speaker (no fixed speaker count) diarization for long meetings, built on:

  • NVIDIA Nemotron-3-Diarization slot head — frame-level slot posteriors (8 slots)
  • WeSpeaker ResNet34-LM — 256-d speaker embeddings
  • official e-means (Meta's faiss contrib, vendored MIT) — clusters embeddings within slots, creating new speakers on demand

Upload a meeting clip and get back who spoke when: a color-coded speaker-turn timeline, a per-speaker summary, and standard RTTM output.

  • streaming mode (default): causal — audio is fed in 100 ms chunks as a live system would; a bounded 90 s window is re-clustered every chunk, labels may be revised for 10 s, and speaker ids are anchored to slots via Hungarian matching.
  • offline mode: the whole recording at once — best accuracy, not causal.

The diarization code is Kyudan/eDIAR (Apache-2.0), vendored in emeans_diar/. Model weights are downloaded at runtime from their own repos: nvidia/Nemotron-3-Diarization (openmdw-1.1) and Wespeaker/wespeaker-voxceleb-resnet34-LM (CC-BY-4.0).

Example audio: synthetic LibriSpeech conversations from the eDIAR project page (LibriSpeech is CC-BY 4.0).

Clips are capped at 3 minutes in this demo — the method itself targets hour-long meetings and large archives.