SASVC

Spoofing-Aware Speaker Verification More information can be found in the webpage. This project provide three ways to realize Sooofing-aware Speaker Verification system. Score Fusion, embedding fusion and muti-task learning.

Prepare

1. Prepare embedding file

You can refer this offical [project](https://github.com/search?q=SASVC2022). Download the Embedding file. Put the embeddings floder to SF and EF folder.

2. Tools

This project need install this tools first
1. [kaldi](https://github.com/kaldi-asr/kaldi)
2. [ASV-subtools](https://github.com/Snowdar/asv-subtools)
3. [speechbrain](https://github.com/search?q=speechbrain)

Score Fusion

$$S_{sasv} = S_{cm} * sigmoid（S_{asv}）$$

ASV: ECAPA-TDNN + PLDA + LIP-Reg Adaptation SV-EER 1.47%

It takes PLDA backend as classifier, in this way. Unsupervised domain adaptation and supervised domain adaptation were applied in score fusion method to improve the speaker verification performance.

CM: wav2vev-ASSIST(Pretrain + Finetune CM-EER 0.20%)

score produced by Wav2Vec-AASIST EER was 0.20%. The sasv score is the multiplication of asv score and cm score processed by sigmoid function.

$ cd SF && ./run.sh

Embedding Fusion

Using Conv1D layers and SEModule to train a SASV model using embeddings from pre-trained asv system and countermeature system provided by SASV baseline system.

Training it:

cd EF && python main.py --config ./configs/sasv.conf

Multi-task

In this method, i used a pretrained asv system and Attentive statistic pooling layers and fusion backend to build a SASV model. In this way, speechbrain toolkit is needed.

Train:

cd EF && python train.py yaml/sasv.yaml

Test:

python eval.py yaml/sasv.yaml

Results(Eval set):

Model	SASV(EER%)	SV(EER%)	SPF(EER%)
Score Fusion	1.06	1.53	0.64
Embedding Fusion	0.96	1.24	0.68
Multi-task	3.24	3.99	1.64