DPT-FCNET: Conformer-based Dual-path Network for Monaural Speech Enhancement
Zhongchao Guan, Zihao Feng, Ying Gao, Shifeng Ou · 2024
Monaural speech enhancement has made extensive progress in the field of amplitude. However, since speech signals in practice are mostly in complex-valued form, it is clearly insufficient to enhance the performance of speech in terms of amplitude characteristics. Recently, T-F dual-path based conformer networks as well as expansion module based encoders and decoders have achieved excellent results in the field of speech enhancement. Therefore, we propose a dual-stream dual-decoding conformer network (DPT-FCNet) based on T-F dual-path for monaural speech enhancement. DPT-FCNet applies the T-F dual-path self-attention mechanism to efficiently model the complex-valued sequences while preserving the phase estimation, and the dual-decoding framework with complex masking and complex spectral mapping provides the model with better learning capability. As a result, DPT-FCNet fully demonstrates the advantages of dual-stream dual-path processing and joint dual-decoding training objectives, and shows advanced performance in VoiceBank+Demand benchmark tests.