E2E-SD: Clustering-Free End-to-End Spoof Diarization with Encoder-Decoder Attractors

This repository contains the trained default model of Clustering-Free End-to-End Spoof Diarization with Encoder-Decoder Attractors (NeurIPS 2026).

Code: github.com/Bongsujung/E2E-SD

Model

• best.pt (1.4 GB): WavLM-Large front-end, gMLP back-end and encoder-decoder attractors. We trained it on the PartialSpoof training set (one run, seed 777).
• The checkpoint contains the SSL front-end. Thus, inference does not download other weights.

Usage

git clone https://github.com/Bongsujung/E2E-SD.git && cd E2E-SD
pip install -r requirements.txt -e .
wget -P exp/e2esd/s777 https://huggingface.co/GoutSIDE/E2E-SD/resolve/main/best.pt
python scripts/infer.py --checkpoint exp/e2esd/s777/best.pt --db-root $DB --split eval \
    --out-dir exp/e2esd/s777/eval
python scripts/evaluate.py --infer-dir exp/e2esd/s777/eval --db-root $DB --split eval

$DB is the PartialSpoof database directory. The GitHub README gives the installation and the data preparation.

Results

PartialSpoof evaluation set, estimated K (%). Lower is better, except for K-acc.

JI	JER	Known	Varied	Unknown	K-acc	seg-EER	utt-EER
7.03	14.92	9.18	8.79	20.31	90.35	14.94	1.14

The paper gives the mean over three runs (JER 16.01 ± 0.95).

Citation

@inproceedings{jung2026e2esd,
  title     = {Clustering-Free End-to-End Spoof Diarization with Encoder-Decoder Attractors},
  author    = {Jung, Bongsu and Kim, Donghee and Kim, Wooil},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

License

MIT. The training data (PartialSpoof) and the pre-trained WavLM-Large have their own licenses.