This repository provides the implementation of the paper: Replay and Synthetic Speech Detection with Res2Net architecture.

1. ResNet blocks

2. Overall model architecture

1. System performance on the ASVspoof2019 PA and LA dataset. (The input features for PA and LA are Spec and LFCC, respectively.)

2. System performance on the ASVspoof2019 PA (left) and LA (right) of SE-Res2Net50 with different acoustic features.

1. Python and packages

   This code was tested on Python 3.7 with PyTorch 1.6.0. Other packages can be installed by:

   pip install -r requirements.txt

2. Kaldi

   This work used Kaldi to extract features, you need to install Kaldi before running our scripts.

3. MATLAB

   The LFCC feature adopted in this work is extracted via the MALTAB codes privided by ASVspoof2019 orgnizers.

This work is conducted on ASVspoof2019 Dataset, which can be downloaded via https://datashare.ed.ac.uk/handle/10283/3336. It consists of two subsets, i.e. physical access (PA) for replay attacks and logical access (LA) for synthetic speech attacks.

This repository mainly consists of two parts: (i) feature extraction and (ii) system training and evaluation.

Three features are adopted in this repo, i.e. Spec, LFCC and CQT. The top script for feature extraction is extract_feats.sh, where the first step (Stage 0) is required to prepare dataset before feature extraction. It also provides feature extraction for Spec (Stage 1) and CQT (Stage 2), while for LFCC extraction, you need to run the ./baseline/write_feature_kaldi_PA_LFCC.sh and ./baseline/write_feature_kaldi_LA_LFCC.sh scripts. All features are required to be truncated by the Stage 4 in extract_feats.sh.

Given your dataset directory in extract_feats.sh, you can run any stage (e.g. NUM) in the extract_feats.sh by

./extract_feats.sh --stage NUM

For LFCC extraction, you need to run

./baseline/write_feature_kaldi_LA_LFCC.sh
./baseline/write_feature_kaldi_PA_LFCC.sh

This repo supports different system architectures, as configured in the conf/training_mdl directory. You can specify the system architecture, acoustic features in start.sh, then run the codes below to train and evaluate your models.

./start.sh

Remember to rename your runid in start.sh to differentiate each configuration. From our experiments after ICASSP 2021 submission, we observe that SERes2Net50 configured with 14w_8s and 26w_8s can achieve slightly better performance.

For evaluating systems, you can either use the Kaldi command compute-eer with the resulting *.eer file to compute system EER, e.g.

. ./path.sh
compute-eer NameofScoringFile.txt.eer

or use the ASVspoof2019 official script scoring/evaluate_tDCF_asvspoof19.py with the resulting *.txt file to compute both system EER and t-DCF, e.g. on the LA evalation set, you need to run

python scoring/evaluate_tDCF_asvspoof19.py scoring/la_asv_scores/ASVspoof2019.LA.asv.eval.gi.trl.scores.txt NameofScoringFile.txt

If this repo is helpful with your research or projects, please kindly star our repo and cite our paper as follows:

@article{li2020replay,
  title={Replay and Synthetic Speech Detection with Res2net Architecture},
  author={Li, Xu and Li, Na and Weng, Chao and Liu, Xunying and Su, Dan and Yu, Dong and Meng, Helen},
  journal={arXiv preprint arXiv:2010.15006},
  year={2020}
}

• Xu Li at the Chinese University of Hong Kong ([email protected], [email protected])
• If you have any questions or suggestions, please feel free to contact Xu Li via [email protected] or [email protected].