A DCDP-Net Speech Enhancement Model for Parallel Denoising of Amplitude and Phase Spectra

In this paper, we propose DCDP-Net for real-time speech enhancement in the time-frequency domain, which denoises both the magnitude and phase spectra. DCDP-Net uses an encoder-decoder architecture connected by a dilated complex dual-path conformer (DCDP-Conformer). The DCDP-Conformer adds temporal and frequency attention to capture speech information in both domains. The encoder processes the noisy input, while the decoder, consisting of a magnitude mask decoder and a phase decoder, recovers the clean speech. Experiments show that DCDP-Net achieves a PESQ of 3.05 on the VoiceBank+DEMAND dataset, outperforming baseline models. Ablation studies confirm the effectiveness of each module.