Research On Ear Speech Enhancement Algorithm Based On Subband Analysis

As a special form of speech,whispered speech is a common way of human communication.Because its pronunciation is similar to noise,it is very difficult to reduce the noise of whispered speech.The whispered speech enhancement technology is very important in daily life.In this paper,the characteristics of whispered speech in time and frequency domain are analyzed,and then the whispered speech enhancement technology is studied.Since whispered speech is also a kind of speech,before studying whispered speech enhancement,this paper makes a deep study of traditional speech enhancement technology,and proposes a single channel speech enhancement algorithm based on sub-band spectral entropy.Then two improved algorithms are presented based on the research of whispered speech technology in the following paper:the whispered speech enhancement based on improved Mel frequency scale and phase modification as well as the whispered speech enhancement combining the formant characteristics of whispered speech and the improved Mel frequency scale.At present,there are few studies on whispered speech,mostly focusing on the analysis of basic characteristics of whispered speech.In this paper,the characteristics of whispered speech are analyzed from many aspects,including physiological property,acoustic characteristics,time and frequency domain simulation analysis,which provides a theoretical basis for later whispered speech enhancement algorithms.A single channel speech enhancement algorithm based on sub-band spectral entropy is presented in this paper.In the multi-band spectral subtraction method,ERB scale is used to divide the critical frequency band,and the speech spectrum component and noise spectrum component are judged by calculating the sub-band spectral entropy.The noise update threshold is compared with the a posteriori signal-to-noise ratio to estimate the noise spectrum.Finally,the speech signal is enhanced by spectral subtraction alone in each frequency band.The experimental results show that the proposed single channel speech enhancement algorithm outperforms other comparable traditional speech enhancement algorithms.On the basis of whispered speech specificity analysis,a whispered speech enhancement algorithm based on improved Mel frequency scale and phase modification is proposed.Firstly,the new Mel frequency scale is improved according to the traditional Mel frequency scale and the two frequency scales proposed by Xu Boling.In addition,the traditional phase spectrum compensation is modified,and the modified phase spectrum is combined with the reduced whispered speech amplitude spectrum to obtain the final enhanced whispered speech spectrum.The estimated noise spectral spectrum in the phase spectrum compensation function is obtained by a new method.Finally,the experimental performance analysis shows that the improved whispered speech enhancement algorithm outperforms the other two comparable whispered speech enhancement algorithms in objective and subjective evaluation.Moreover,a whispered speech enhancement algorithm based on the combination of the whispered speech formant characteristics and the improved Mel frequency scale is presented.Firstly,according to the improved Mel frequency scale of the previous improved whispered speech enhancement algorithm,the Mel frequency scale that can be selected in the improved algorithm is obtained,and the energy-entropy ratio of the whispered speech is calculated for end point detection.Then,the second formant frequency of the whispered speech is extracted from the segment.According to the frequency,the appropriate Mel frequency scale is selected for multi-band spectral subtraction,and the final enhanced whispered speech signal is obtained.The experimental results show that the improved whispered speech enhancement algorithm excels the other two whispered speech enhancement algorithms in both objective and subjective evaluation.