nesl / asvspoof2019

Our submission to the ASVspoof 2019: Automatic Speaker Verification Spoofing and Countermeasures Challenge
96 stars 40 forks source link

Deep Residual Neural Networks for Audio Spoofing Detection

This repo has an implementation for our paper Deep Residual Neural Networks for Audio Spoofing Detection, this is also describes the solution of team UCLANESL in the ASVSpoof 2019 competition.

Dataset

The ASVSpoof2019 dataset can be downloaded from the following link:

ASVSpoof2019 dataset

Training models

python model_main.py --num_epochs=100 --track=[logical/physical] --features=[spect/mfcc/cqcc]   --lr=0.00005

Please note that the CQCC features are computing using the Matlab code in cqcc_extraction.m, so you need to run this file to generate cache files of CQCC featurs before attempting to traiin or evaluate models with CQCC features.

To perform fusion of multiple results files

 python fuse_result.py --input FILE1 FILE2 FILE3 --output=RESULTS_FILE

Evaluating Models

Run the model on the evaluation dataset to generate a prediction file.

python model_main.py --eval  --eval_output=RESULTS_FILE --model_path=CHECKPOINT_FILE

Then compute the evaluation scores using on the development dataset

python evaluate_tDCF_asvspoof19.py RESULTS_FILE PATH_TO__asv_dev.txt