Introduction

This repository contains Ruijie Tao's unofficial reimplementation of the standard ECAPA-TDNN.

This repository is modified based on voxceleb_trainer and TaoRuijie/ECAPA-TDNN.

In this project, we use CN-Celeb2 Dataset to train ECAPA-TDNN model.

Best Performance in this project

Dataset	CN-Celeb2
EER	2.99%
Accuracy	79.60%

---

Dependencies

Create the environment:

conda create -n ECAPA python=3.7
conda activate ECAPA
pip install -r requirements.txt

Data preparation

Dataset for training:

1. CN-Celeb2 training set;

wget https://us.openslr.org/resources/82/cn-celeb2_v2.tar.gzaa
wget https://us.openslr.org/resources/82/cn-celeb2_v2.tar.gzab
wget https://us.openslr.org/resources/82/cn-celeb2_v2.tar.gzac
cat cn-celeb2_v2.tar.gzaa cn-celeb2_v2.tar.gzab cn-celeb2_v2.tar.gzac > cn-celeb2_v2.tar.gz
tar -xzvf cn-celeb2_v2.tar.gz

2. MUSAN dataset;

wget https://us.openslr.org/resources/17/musan.tar.gz
tar -xzvf musan.tar.gz

3. RIR dataset.

wget https://us.openslr.org/resources/28/rirs_noises.zip
unzip rirs_noises.zip

Training

We provide data/train_files.txt and data/test_pairs.txt. You can use these lists to train and evaluate.

Or you can use GenerateList() in tools.py to generate the train_list.txt and test_pairs.txt randomly.

Then you can change the data path in the train.py. Train ECAPA-TDNN model end-to-end by using:

python train.py --save_path exps/[your_exp_name]

Every test_step epoches, system will be evaluated and print the EER.

The result will be saved in exps/[your_exp_name]/score.txt. The model will saved in exps/[your_exp_name]/model

In my case, I trained 80 epoches in one 3090 GPU. Each epoch takes 90 mins.

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}
}

Ruijie Tao's 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}
}

Acknowledge

We study many useful projects in our codeing process, which includes:

clovaai/voxceleb_trainer.

lawlict/ECAPA-TDNN.

speechbrain/speechbrain

ranchlai/speaker-verification

Thanks for these authors to open source their code!