Speech dereverberation method with convolutional neural network and reverberation time attention

A reverberation suppression algorithm based on convolutional neural network-based encoder-decoder with reverberation time attention mechanism is proposed.The algorithm achieves reverberation suppression through the encoder-decoder model and uses the reverberation time attention mechanism to overcome the effect of different reverberation environments for reverberation suppression performance.In the encoder,the convolutional kernels with different sizes are applied to the reverberant magnitude spectrum to encode the features with multi-scale context information.The attention module is introduced to selectively focus on the encoded features to generate weighted feature under the different reverberation times.The magnitude spectrum of the dereverberated signal is finally reconstructed using the weighted feature in the decoding process.In simulated and real reverberation environments,our proposed method has achieved 0.36 dB and 0.66 dB improvements in the speech reverberation modulation energy ratio compared to the baseline system.Experimental results show that our proposed algorithm can adapt to various reverberation environments and has higher robustness in real reverberation environments.