GD-Conformer: a Conformer-based gated dense encoder-decoder for monaural speech enhancement

Abstract Speech enhancement improves speech quality by mitigating noise, dereverberation, and echo. Existing methods struggle with amplitude-phase compensation, capturing temporal-frequency features, and high complexity. To address these issues, a gated dense encoder-decoder architecture with a two-stage Conformer, abbreviated as GD-Conformer, is proposed. It integrates a gated dense module, a two-stage residual Conformer module, a mask decoder and a complex decoder. The gated dense module consists of two parts: a dilated dense convolution and a gated convolution, where the former captures both global and local dependencies features, while the latter refines these distinct features accordingly. The two-stage residual Conformer focuses on the time-frequency dependence of speech, it also reduces the computational complexity. The mask decoder and the complex decoder restore spectral resolution while preserving speech fidelity. The outcomes of experiments conducted on the public dataset VoiceBank+DEMAND and DNS Challenge 2020 demonstrate that, compared with those state-of-the-art methods, the proposed GD-Conformer achieves comparable performance in terms of denoising and generalization with fewer parameters and lower computation complexity.