Dual-path and interactive UNET for speech enhancement with multi-order fractional features

Abstract

Preprocessing techniques for denoising and enhancement play a crucial role in significantly improving speech recognition performance. In neural-network-based speech enhancement methods, input features provide the network with essential information to learn from the data. In this study, we introduced multi-order fractional features into a speech enhancement network. These features can represent fine details and offer the advantages of multidomain joint analysis, thereby expanding the input information available to the network. Subsequently, a new dual-path UNET network was designed, in which pure speech and noise are estimated separately. By leveraging the complementarity of the two-branch target estimation, we introduced a fractional information interaction module between the two paths for parameter optimization. Finally, the association module combined the two output information streams to enhance the speech performance. The results from ablation experiments demonstrated the effectiveness of both the multi-order fractional features and the improved dual-path network. Comparison experiments revealed that the proposed algorithm significantly improved speech quality and intelligibility.