Mdct-Dpanet: Dual-Path Attention Network for Multi-Channel Speech Separation

Speech separation algorithm is designed to extract the target speech from mixed audio signals and has wide application in complex acoustic environment, such as noisy and multi-speaker conversations. This paper integrates the dual-path structure with various attention mechanisms and proposes a low-computational-complexity multi-channel speech separation algorithm called MDCT-DPANet (Modified Discrete Cosine Transform Dual-Path Attention Network). Experimental results on fixed-topology microphone array speech separation demonstrate that MDCT-DPANet achieves superior performance while maintaining a relatively small model size and computational complexity. The proposed architecture exhibits significant advantages in balancing separation accuracy with computational efficiency compared to conventional approaches, making it particularly suitable for resource-constrained scenarios.