TD-TFDeReNet: A Hybrid Framework for Robust Speech Dereverberation Using Time-Domain and Time-Frequency Attention Mechanisms
Bin Liu, Baosheng Lv, Yu Zhao, Yi Xin Zhou · 2025
Reverberation in complex acoustic environments presents substantial challenges for audio conferencing systems, impairing speech quality and system robustness. To address this issue, a novel framework named Time-Domain and Time-Frequency Dereverberation Network (TD- TFDeReN et) is proposed. This framework integrates a time-domain convolutional network and a time-frequency attention mechanism to perform speech dereverberation. The model captures short-term temporal features and long-term spectral dependencies, facilitating the recovery of clean speech from reverberant inputs. Experimental results demonstrate that TD-TFDeReN et outperforms the baseline models, including U -Net and LS U-Net, across multiple objective metrics. This approach enhances speech quality and shows strong potential for real-world applications.