Rawformer Implementation for Anti-Spoofing

This is my own implementation of Rawformer model (LEVERAGING POSITIONAL-RELATED LOCAL-GLOBAL DEPENDENCY FOR SYNTHETIC SPEECH DETECTION - Xiaohui Liu, Meng Liu, Longbiao Wang, Kong Aik Lee2, Hanyi Zhang1, Jianwu Dang)

WARNING

• This code could be not same as explained in the paper. -If you find some bugs and want to fix it, please open pull requests.
• Using pre-emphasis for preprocessing showed superior performance.

Rawformer-S vs Rawformer-L vs SE-Rawformer

In the paper, authors developed three types of Rawformer, Rawformer-S, Rawformer-L and SE-Rawformer. I implemented all of these models only with 1-dimesional positional encoding. N is the number of Conv2D-based Blocks and M is the number of Transformer Encoders.

• Rawformer-S
  • N = 4
  • M = 2
  • Conv2D-based Block - same as a ResNet block used in AASIST
• Rawformer-L
  • N = 6
  • M = 3
  • Conv2D-based Block - same as a ResNet block used in AASIST
• SE-Rawformer
  • N = 4
  • M = 2
  • Conv2D-based Block - replaced blocks of Rawformer-S with Res-SERes2Net blocks for last three blocks