ASVspoof2021_AIR

UR Channel-Robust Synthetic Speech Detection System for ASVspoof 2021

This repository contains our implementation of the paper, "UR Channel-Robust Synthetic Speech Detection System for ASVspoof 2021". Paper link here Video link here

The code here is a little bit unorganized, please feel free to open any issues. The contribution from this work mainly derives from our previous research works. Please feel free to check out the paper and code there.

Reference to our prior research works

[1] Y. Zhang, F. Jiang and Z. Duan, "One-Class Learning Towards Synthetic Voice Spoofing Detection," in IEEE Signal Processing Letters, vol. 28, pp. 937-941, 2021, doi: 10.1109/LSP.2021.3076358. [link] [arXiv] [code] [video]

[2] Y. Zhang, G. Zhu, F. Jiang, Z. Duan, "An Empirical Study on Channel Effects for Synthetic Voice Spoofing Countermeasure Systems," Proc. Interspeech 2021, pp. 4309-4313, 2021, doi: 10.21437/Interspeech.2021-1820. [link] [arXiv] [code] [video]

Requirements

python==3.6

pytorch==1.1.0

Data Channel Augmentation

We use impulse responses from this repo and the code we adapted is in the channel_simulation folder. Please refer to the paper for the augmentation process.

Data Preprocess

The raw wave files of the ASVspoof 2019 dataset is stored in /data/neil/DS_10283_3336. The raw_dataset.py is a pytorch dataloader for the original speech dataset. The feature_extraction.py is to write feature extraction methods, currently I only have LFCC calculation. You may include CQCC, melspectrogram, CQT, MFCC, etc. in the future for model fusion. The utils_dsp.py are some helper functions for feature extraction. The process.py is using the functions in feature extraction and output the preprocessed speech dataset.

Data Preparation

The dataset.py is a pytorch dataloader for the preprocessed speech dataset. Make sure you change the directory path to the path on your machine.

Model and loss functions

The model.py include some model architectures. The loss.py defines some loss functions, including what we proposed.

Run the training code

In the main_train.py, please read the argparse to get familiar with the options. Before running the main_train.py, please change the path_to_features according to the files' location on your machine.

python3 main_train.py --add_loss ang_iso -o /data/xinhui/models/ocsoftmax --gpu 3

Run the test code with trained model

I included the test code in generate_score.py. The eval_metrics.py and evaluate_tDCF_asvspoof19.py are some helper functions to calculate the evaluation metrics. Before running, please change the model_dir to the location of the model you would like to test with.

python3 generate_score.py

Citations

@inproceedings{chen21_asvspoof,
  author={Xinhui Chen and You Zhang and Ge Zhu and Zhiyao Duan},
  title={{UR Channel-Robust Synthetic Speech Detection System for ASVspoof 2021}},
  year=2021,
  booktitle={Proc. 2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge},
  pages={75--82},
  doi={10.21437/ASVSPOOF.2021-12}
}