Parallel-Path Transformer Network for Time-Domain Monaural Speech Separation

In recent years, speaker-independent, monaural speech separation methods have made great progress with the development of deep neural networks (DNNs). However, in the automotive environment, computational resources are limited and the cockpit sound field environment is more complex. These factors pose significant challenges to the task of speech separation. To address these challenges, this paper proposes a parallel-path transformer model. The model employs a parallel processing strategy that combines improved feed-forward networks with transformer modules. It simultaneously applies intra-chunk transformer and inter-chunk transformer to process the input sequences, avoiding the need for implicit modeling of intermediate states. This approach enables the model to perform local and global modeling of the speech signals in a parallel manner, capturing both short and long-term dependencies within the speech sequences. Consequently, the proposed model enhances the system’s modeling performance. Experimental results on the WSJ0-2Mix and WHAM! datasets demonstrate that the proposed model achieves excellent speech separation performance while maintaining smaller model parameters and computational complexity.