Skip to content

alexw914/SASVC

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

16 Commits
 
 
 
 
 
 
 
 

Repository files navigation

SASVC

Spoofing-Aware Speaker Verification

This project provide three ways to realize Sooofing-aware Speaker Verification system. Score Fusion, embedding fusion and muti-task learning.

Score Fusion

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. It needs kaldi and asv-subtools. When using pretrained model of ECAPA-TDNN and LIP-Reg adaptation, it gets Speaker Verification EER 1.47%. Countermeasure score produced by Wav2Vec-AASIST EER was 0.20%. The sasv score is the sum of asv score cm score produced by sigmoid function. The finald results on eval set is SASV: 1.06%, SV: 1.53%, SPF: 0.64%.

Embedding Fusion

Using Conv1D layer and SEModule to train a SASV model using embeddings from pre-trained asv system and countermeature system. It get results of SASV: 0.96% SV: 1.24% SPF: 0.68% on SASV eval set.

Multi-task

In this method, i used a pretrained asv system and Attentive statistic pooling layers and backend to build a SASV model. In this way, speechbrain toolkit is needed. SASV: 3.24% SV: 3.99% SPF: 1.64% on eval set.

About

Spoofing-Aware Speaker Verification

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published