# SASV Challenge 2022
- https://arxiv.org/pdf/2201.10283.pdf
- https://sasv-challenge.github.io/
- https://github.com/sasv-challenge/SASVC2022_Baseline

## Idea
- Use the three embeddings (enroll embedding A, testing embedding B, cm embedding C) as query, key and values (using 2 transformers, one take A, B, C as query, key, value, and the other take A, B, C as query, value, key) to train a scoring model

## reference
### https://arxiv.org/pdf/2006.05599.pdf
- use the anti-spoofing label to help asv training

### https://arxiv.org/pdf/2202.05253.pdf
- use asv score to help the anti-spoofing layer training. Large asv score with result in large gradient (if the trial is spoofing, then it will benefit from the large gradient).

## Experiments