Research On Speech Enhancement Based On Noise Bases And Its Robustness

Speech enhancement is needed in speech communication process due to the de-creased intelligibility and speech quality by background noise and reverberation.For the speech enhancement based on unsupervised learning,the unreasonable assumption limits its performance.In recent years,with the improvements of data scale and hard-ware performance,the speech enhancement based on deep neural network(DNN)shows the great advantage over unsupervised learning methods.Firstly,we introduce a speech enhancement method based on DNN.But this method can not measure or control the coverage of noise in collection of real noise.In other words,this method only focuses on the size,not detailed analysis of data.Usually the data is redundant when data size is large.In addition,the collection of real noise usu-ally costs too much for the general researcher.To solve the problems above,we propose a DNN speech enhancement method based on noise bases,and a systematic study on noise robustness is proposed.Secondly,taking account of the diversity and compactness of noise,we propose a DNN speech enhancement method based on noise bases.The learning of DNN input and output is at the frame level,which makes it possible for the analysis on smaller unit of the noise spectrum.Therefore,we first verify the noise learning principle of DNN.Using the noise bases rationally constructed instead of real noise,the DNN speech en-hancement method based on noise bases yields comparable performance to the tradi-tional method of using real noise.Furthermore,there is complementarity between the noise bases and real noise.Furthermore,in order to get the speech and noise bases fully mixed,and to cover more types of real noise with noise bases when data size is limited,we propose a method of constructing noise signals based on noise bases and linear combination.We first introduce the principle that noise bases cover more noise types using linear combination.Then the performance of the noise bases is further improved by using the noise signal based on the noise bases and linear combination.When customized for narrowband noise,this method yields better performance in unseen narrowband noise than that using 50 real noise types.In addition,the training efficiency is doubled.Finally,using the new speech enhancement frameworks based on progressive learn-ing and multi-information source fusion,we verify our conclusions with multiple types of narrow band and wide band noise and speech out of training set.That is,the noise bases yields comparable performance to the traditional method of using real noise in the new frameworks,which does not require any real noise and shows the generalization ability to the multiple type of noise out of training set.