GLASS: Investigating Global and Local context Awareness in Speech Separation

Previous speech separation systems commonly employ the Dual-Path (DP) mechanism. The DP mechanism addresses optimization challenges posed by considerable sequential input lengths, yet its compulsory interleaving pattern for local and global feature extraction raises concerns regarding optimal utilization of features across different layers. This study emphasizes the need for parallel processing of global and local information in speech separation, proposing the Global and Local context-Aware Speech Separation method (GLASS). GLASS integrates self-attention and convolutional layers into a parallel design, demonstrating state-of-the-art performance in both anechoic and noisy settings. The findings reveal patterns in the relevance of local and global information across layers, underscoring the significance of proper architecture in improving speech separation systems.