The Direction-of-Arrival Combined Dual-channel Unsupervised Speech Enhancement Technology

Speech enhancement can improve the signal quality of microphones in the noisy environment.After enhancement,the signal will be more precise,which not only brings the listener a more comfortable sense of hearing but also makes it easier for auto speech recognition.Unsupervised speech enhancement methods are more robust in different application scenarios lacking prior knowledge.The article is mainly divided into the following three parts:First,the framework of the mask-based beamforming system to be used in this article is introduced,including dereverberation and various beamforming methods.This framework will serve as the basis for the following article to discuss the influence of different masking estimation methods on speech enhancement.Secondly,the method of the unsupervised probability model for mask estimation is introduced.The unsupervised probability model includes complex Watson mixture model,complex Bingham mixture model,complex angle center Gaussian mixture model,and complex Gaussian mixture model.These several methods can effectively realize the enhancement of the speech signal.When iterating only once,the four models have similar performance.At the beginning of the iteration,the estimated masks show similar for random initialization.As the number of iterations increases,the model gradually converges,and the estimated mask is more accurate.Finally,direction-guided unsupervised speech enhancement method is introduced.In this method,the direction of arrival of the target speaker is estimated and then used to construct the spatial covariance matrix,which serves as the initialization of the model parameters.This method can be used in the complex angle center Gaussian mixture model,complex Gaussian mixture model,and the complex Watson mixture model.In the experiments,the results of combining the direction of arrival are significantly better than those of random initialization,especially in the iterative when the number of iterations is low.The result shows that directional information improves the speech enhancement performance,reduces the WER by 4.33%,and improves the SDR by 2.60 dB the most.