ASSERT: Anti-Spoofing with Squeeze-Excitation and Residual neTworks

This repository contains codes to reproduce the core results from our Interspeech 2019 paper: ASSERT: Anti-Spoofing with Squeeze-Excitation and Residual neTworks

If you find the code useful, please cite

@article{lai2019assert,
  title={ASSERT: Anti-Spoofing with Squeeze-Excitation and Residual neTworks},
  author={Lai, Cheng-I and Chen, Nanxin and Villalba, Jes{\'u}s and Dehak, Najim},
  journal={arXiv preprint arXiv:1904.01120},
  year={2019}
}

Abstract

We present JHU's system submission to the ASVspoof 2019 Challenge: Anti-Spoofing with Squeeze-Excitation and Residual neTworks (ASSERT). Anti-spoofing has gathered more and more attention since the inauguration of the ASVspoof Challenges, and ASVspoof 2019 dedicates to address attacks from all three major types: text-to-speech, voice conversion, and replay. Built upon previous research work on Deep Neural Network (DNN), ASSERT is a pipeline for DNN-based approach to anti-spoofing. ASSERT has four components: feature engineering, DNN models, network optimization and system combination, where the DNN models are variants of squeeze-excitation and residual networks. We conducted an ablation study of the effectiveness of each component on the ASVspoof 2019 corpus, and experimental results showed that ASSERT obtained more than 93% and 17% relative improvements over the baseline systems in the two sub-challenges in ASVspooof 2019, ranking ASSERT one of the top performing systems.

Single Model Benchmark Results

Note: Evaluation key is not released yet, so we only present dev results below.

Download Pretrained Model Weights

We included the pretrained model weights. We do not plan to release the pretrained models for Mean-Std ResNets.

• SENet34: ./pretrained/pa/senet34 and ./pretrained/la/senet34
• SENet50: ./pretrained/pa/senet50 and ./pretrained/la/senet50
• Attentive Filtering Network: ./pretrained/pa/attentive_filtering_network and ./pretrained/la/attentive_filtering_network
• Dilated ResNet: ./pretrained/pa/dilated_resnet and ./pretrained/la/dilated_resnet

Dependencies

This project uses Python 2.7. Before running the code, you have to install

• PyTorch 0.4
• Sacred
• Kaldi

The former 2 dependencies can be installed using pip by running

pip install -r requirements.txt

Getting Started

./assert/ contains the main source codes, ./baseline/ contains the code for CQCC-GMM, LFCC-GMM and i-vectors, and ./features/ contains acoustic feature extraction codes.

Make sure to read through the ASVspoof 2019 webpage and download the dataset.

Baselines

./baseline/ contains code for the CQCC-GMM, LFCC-GMM, and i-vectors baselines. ./baseline/baseline_CM.m is the official MATLAB script for CQCC-GMM and LFCC-GMM. Make sure to organize the dataset according to the comments within the script. To train the baselines, do

./baseline_CM.sh

Features

./features/ contains code for acoustic feature extraction. Follow ./features/run_feature.sh stage by stage for acoustic feature extraction. Be sure to set up Kaldi and modify ./features/path.sh to point to your own Kaldi directory. To extract features, do

./run_feature.sh.sh

ASSERT

./assert/ contains the code for ASSERT. ./assert/main.py contains the model training script and model hyperparameters, and ./assert/src/ contains model implementations. To train the models, do

python main.py

To trian the model on GPU (recommended), do

CUDA_VISIBLE_DEVICES=`free-gpu` python main.py

Experiments will be saved automatically by Sacred to ./assert/snapshots/

Authors

Cheng-I Lai, Nanxin Chen, Jesús Villalba, Najim Dehak

If you encouter any problem, feel free to contact me.