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.

Model	PA dev min-tDCF	PA dev EER (%)	LA dev min-tDCF	LA dev EER (%)
SENet34	0.01514	0.5751	0.0	0.0
SENet50	0.01709	0.6317	0.0	0.0
Attentive Filtering Network	0.02096	0.7407	0.0	0.0
Dilated ResNet	0.02377	0.7798	0.0	0.0
Mean-Std ResNet	0.022	0.832	0.0	0.0
CQCC-GMM	0.195	9.87	0.012	0.43
LFCC-GMM	0.255	11.96	0.066	2.71
100-i-vectors	0.306	12.37	0.155	5.18
200-i-vectors	0.322	12.52	0.121	4.12

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.