Dual channel neural network speech enhancement algorithm based on time frequency masking

In order to improve the ability of speech enhancement algorithm eliminate directional noise and suppress reverberation, combining the advantages of single and multi-channel processing signals, a dual-channel neural network time-frequency masking speech enhancement algorithm was proposed. First, using the improved multi-resolution cochlear dynamic and static features (DSMRACC), combined with an adaptive mask (AM) optimized based on the signal-to-noise ratio (SNR), the dual-microphone signals were separately enhanced by a single channel deep neural network (DNN) to achieve the goal of fully utilizing the nonlinear features of speech to improve perception. Second, a steering vector localization method based on the AM was proposed to accurately calculate the spatial covariance matrix and steering vectors, locate the target speech accurately under the noise and reverberation environment. Finally, signal was input to a convolutional beamformer to further denoise and suppress reverberation. The experimental results show that compared with other speech enhancement algorithms, the enhanced speech has better speech quality and intelligibility. © 2021 Editorial Board of Journal of Huazhong University of Science and Technology. All right reserved.