Wespeaker implementation

I re-wrote the whole system with wespeaker toolkit and achieved higher results, which can be found here

Model	AS-Norm	LMFT	QMF	vox1-O-clean	vox1-E-clean	vox1-H-clean
WavLM Base Plus + MHFA	√	×	×	0.750	0.716	1.442
WavLM Large + MHFA	√	×	×	0.649	0.610	1.235

SLT22_MultiHead-Factorized-Attentive-Pooling

This repository contains the Pytorch code of our paper titled as An attention-based backend allowing efficient fine-tuning of transformer models for speaker verification. This implementation is based on vox_trainer.

---

Dependencies & Data preparation

• Pls follow Data preparation for voxceleb, musan, and RIR datasets.
• Download the pre-trained model (e.g. WavLM Base+) from huggingface or its official website.
• Modify the corresponding train_list, test_list, train_path, test_path, musan_path, rir_path in trainSpeakerNet.py and trainSpeakerNet_Eval.py, respectively.
• Add absolute path to the pre-trained model to pretrained_model_path in both yaml/Baseline.yaml and yaml/Baseline_lm.yaml

Training and Testing

name=Baseline.yaml
name_lm=Baseline_lm.yaml

python3 trainSpeakerNet.py --config yaml/$name --distributed >> log/$name.log
python3 trainSpeakerNet.py --config yaml/$name_lm --distributed >> log/$name_lm.log
python3 trainSpeakerNet_Eval.py --config yaml/$name_lm  --eval >> log/$name_lm.log

where lm denotes large margin fine-tuning.

Citation

If you used this code, please kindly consider citing the following paper:

@INPROCEEDINGS{10022775,
  author={Peng, Junyi and Plchot, Oldřich and Stafylakis, Themos and Mošner, Ladislav and Burget, Lukáš and Černocký, Jan},
  booktitle={2022 IEEE Spoken Language Technology Workshop (SLT)}, 
  title={An Attention-Based Backend Allowing Efficient Fine-Tuning of Transformer Models for Speaker Verification}, 
  year={2023},
  volume={},
  number={},
  pages={555-562},
  doi={10.1109/SLT54892.2023.10022775}}