This repository contains my unofficial reimplementation of the standard ECAPA-TDNN, which is the speaker recognition in VoxCeleb2 dataset.
This repository is modified based on voxceleb_trainer.
| Dataset | Vox1_O | Vox1_E | Vox1_H |
|---|---|---|---|
| EER | 0.86 | 1.18 | 2.17 |
| minDCF | 0.0686 | 0.0765 | 0.1295 |
Notice, this result is in the Vox1_O clean list, for Vox1_O Noise list: EER is 1.00 and minDCF is 0.0713.
I have uploaded the system description, please check the Session 3, ECAPA-TDNN SYSTEM.
Note: That is the setting based on my device, you can modify the torch and torchaudio version based on your device.
Start from building the environment
conda create -n ECAPA python=3.7.9 anaconda
conda activate ECAPA
pip install -r requirements.txt
Start from the existing environment
pip install -r requirements.txt
Please follow the official code to perpare your VoxCeleb2 dataset from the 'Data preparation' part in this repository.
Dataset for training usage:
-
VoxCeleb2 training set;
-
MUSAN dataset;
-
RIR dataset.
Dataset for evaluation:
Then you can change the data path in the trainECAPAModel.py. Train ECAPA-TDNN model end-to-end by using:
python trainECAPAModel.py --save_path exps/exp1
Every test_step epoches, system will be evaluated in Vox1_O set and print the EER.
The result will be saved in exps/exp1/score.txt. The model will saved in exps/exp1/model
In my case, I trained 80 epoches in one 3090 GPU. Each epoch takes 37 mins, the total training time is about 48 hours.
Our pretrained model performs EER: 0.96 in Vox1_O set without AS-norm, you can check it by using:
python trainECAPAModel.py --eval --initial_model exps/pretrain.model
With AS-norm, this system performs EER: 0.86. We will not update this code recently since no enough time for this work. I suggest you the following paper if you want to add AS-norm or other norm methods:
Matejka, Pavel, et al. "Analysis of Score Normalization in Multilingual Speaker Recognition." INTERSPEECH. 2017.
We also update the score.txt file in exps/pretrain_score.txt, it contains the training loss, training acc and EER in Vox1_O in each epoch for your reference.
Original ECAPA-TDNN paper
@inproceedings{desplanques2020ecapa,
title={{ECAPA-TDNN: Emphasized Channel Attention, propagation and aggregation in TDNN based speaker verification}},
author={Desplanques, Brecht and Thienpondt, Jenthe and Demuynck, Kris},
booktitle={Interspeech 2020},
pages={3830--3834},
year={2020}
}
Our reimplement report
@article{das2021hlt,
title={HLT-NUS SUBMISSION FOR 2020 NIST Conversational Telephone Speech SRE},
author={Das, Rohan Kumar and Tao, Ruijie and Li, Haizhou},
journal={arXiv preprint arXiv:2111.06671},
year={2021}
}
VoxCeleb_trainer paper
@inproceedings{chung2020in,
title={In defence of metric learning for speaker recognition},
author={Chung, Joon Son and Huh, Jaesung and Mun, Seongkyu and Lee, Minjae and Heo, Hee Soo and Choe, Soyeon and Ham, Chiheon and Jung, Sunghwan and Lee, Bong-Jin and Han, Icksang},
booktitle={Interspeech},
year={2020}
}
We study many useful projects in our codeing process, which includes:
Thanks for these authors to open source their code!
If you meet the problems about this repository, Please ask me from the 'issue' part in Github (using English) instead of sending the messages to me from bilibili, so others can also benifit from it. Thanks for your understanding!
If you improve the result based on this repository by some methods, please let me know. Thanks!
If you are interested to work on this topic and have some ideas to implement, I am glad to collaborate and contribute with my experiences & knowlegde in this topic. Please contact me with ruijie.tao@u.nus.edu.