Overview

This repository is an implementation of the End-to-End Spoof-Aggregated Spoofing-Aware Speaker Verification System. The model performance is tested on the ASVSpoof 2019 Dataset.

Setup

Environment

Show details

speechbrain==0.5.7
pandas
torch==1.9.1
torchaudio==0.9.1
nnAudio==0.2.6
ptflops==0.6.6

Create a conda environment with conda env create -f environment.yml.
Activate the conda environment with conda activate .

Data preprocessing

.
├── data                       
│   │
│   ├── PA                  
│   │   └── ...
│   └── LA           
│       ├── ASVspoof2019_LA_asv_protocols
│       ├── ASVspoof2019_LA_asv_scores
│       ├── ASVspoof2019_LA_cm_protocols
│       ├── ASVspoof2019_LA_train
│       ├── ASVspoof2019_LA_dev
│       
│
└── ARawNet

Download dataset. Our experiment is trained on the Logical access (LA) scenario of the ASVspoof 2019 dataset. Dataset can be downloaded here.
Unzip and save the data to a folder data in the same directory as ARawNet as shown in below.
Run python preprocess_SASASV.py Or you can use our processed data directly under "/processed_data".

Train

python train_SASASV.py yaml/SASASV.yaml --data_parallel_backend -data_parallel_count=2

Evaluate

python eval_SASASV.py

Cite Our Paper

If you use this repository, please consider citing:

@article{teng2022sa, title={SA-SASV: An End-to-End Spoof-Aggregated Spoofing-Aware Speaker Verification System}, author={Teng, Zhongwei and Fu, Quchen and White, Jules and Powell, Maria E and Schmidt, Douglas C}, journal={arXiv preprint arXiv:2203.06517}, year={2022} }