Speech Enhancement Based on Dual-Path Cross Parallel Transformer Network

In this study, a two-path cross parallel transformer neural network is proposed for simultaneous simulation of amplitude and complex spectrum speech enhancement. An innovative structure is introduced, incorporating a residual connected cross parallel transformer module between the encoder and decoder to achieve outstanding performance. Each module can parallelly utilize local and global transformers to extract local and global information, and apply transformers with cross-attention mechanisms to obtain accurate contextual information. To further enhance accuracy, a mask module is introduced for improving gradient propagation. The output of the mask module passes through the attention sensing fusion module, aiming to fuse the amplitude and complex features learned from the dual path structure for optimal spectrum estimation. Through experiments, it has been proven that our dual path cross parallel transformer neural network performs excellently in terms of performance, surpassing other models. In addition, we also conducted ablation experiments to verify the effectiveness of the cross-transformer module and attention perception fusion module. These experiments further support the superiority and innovation of our proposed method.