SE-DSICTWCN: Dual-Stream Information Communication Transformer with CVOCA Nets for Speech Enhancement

Recent advances in Transformer-based models have greatly improved speech enhancement (SE), yet effective phase modeling remains challenging. We propose the Dual-Stream Information Communication Transformer with Convolutional Variational Output Channel Attention Networks (DSICTWCN), a dual-stream encoder-decoder framework with multi-scale CNNs and residual connections for robust time-frequency feature extraction. The architecture incorporates an Information Communication Unit (InforComu) for adaptive magnitude-phase interaction and a Convolutional Variational Output Channel Attention (CVOCA) module for efficient and robust channel-wise feature selection. Four cascaded Dual-Stream Transformer with CVOCA (DSTWC) modules alternately model sub-band and full-band dependencies. Experiments on the VoiceBank+DEMAND benchmark show that DSICTWCN achieves substantial improvements in both speech quality and intelligibility while maintaining a low parameter count, demonstrating its effectiveness for practical SE tasks.