Dual-Path Federative Magnitude and Phase Estimation Combing Transformer with RNN for Single Channel Speech Enhancement

To enhance the effectiveness and adaptability of existing speech enhancement techniques that rely on masking and spectral mapping, especially in complex environments, we propose a novel single-channel speech enhancement approach that integrates multi-order dual-branch magnitude and phase estimation, called DRP-SENet. This method features an encoder-dual-branch decoder architecture. An intermediate layer with C-DPRNN is designed to improve the capture of local patterns in speech spectrograms and dependencies between consecutive frames. Additionally, to obtain voice channel dimension information, Channel-Aware Multi-Head Attention (CH-MHSA) is used in place of Multi-Head Self-Attention (MHSA). Results on the VoiceBank+DEMAND dataset confirm that our proposed method achieves superior subjective and objective metrics compared to most state-of-the-art related methods.