MFFN: Multi-level Feature Fusion Network for monaural speech separation

Abstract

Monaural speech separation based on Dual-path networks has recently been widely developed due to their outstanding processing ability for long feature sequences. However, these methods often exploit a fixed receptive field during feature learning, which hardly captures feature information at different scales and thus restricts the model’s performance. This paper proposes a novel Multi-level Feature Fusion Network (MFFN) to facilitate dual-path networks for monaural speech separation by capturing multi-scale information. The MFFN integrates information of different scales from long sequences by using a multi-scale sampling strategy and employs Squeeze-and-Excitation blocks in parallel to extract features along the channel and temporal dimensions. Moreover, we introduce a collaborative attention mechanism to fuse feature information across different levels, further improving the model’s representation capability. Finally, we conduct extensive experiments on noise-free datasets, WSJ0-2mix and Libri2mix, and the noisy datasets, WHAM! and WHAMR!. The results demonstrate that our MFFN outperforms some current methods without using data augmentation technologies.