Study On Speech Enhancement Algorithm Based On Cochlea Filter And Harmonicity

Speech enhancement means the improvement in intelligibility or quality of a degraded speech signal by using signal processing tools,which is applied in the areas of hand-free communications,speech recognition,car phones,answering machines and so on,and it is a guarantee of high-quality voice communications.However,all of the existing speech enhancement algorithms, with unsatisfactory effect,or high computational complexity,or without noise-enviro- -nment transplantability, are ineffective and difficult to meet the needs of practical application,so it is necessary to study a speech enhancement algorithm to meet the actual voical communications.Systematic classification and summary to existing speech enhancement algorithms are made in this paper,and the existing pitch estimation methods and a functional cochlea simulation model are introduced in detail.lt is obtained that the harmonics are correlated in some degree and the bandwidth of the harmonics are almost constant among different vowels and different persons from the deep thinking of the cochlea simulation model by studying on the features of harmonics of vowels in frequency filed in the way of analyzing a large number of experiments,and according to the two features,a new speech enhancement algorithm—harmonic correlation denoising is proposed.In the proposed method,a typical harmonic is extracted from a certain clean vowel,and then a harmonic correlation filed is constructed by calculating the correlation with the typical harmonic and the noisy speech,furthermore,a large spectrum is constructed based on the harmonic filed and the frequency filed,and the peak points of higher harmonics in the frequency filed are extracted in the large spectrum filed and harmonic correlation filed based on the average threshold method,and then a gammatone filter bank is constructed to filter the noisy speech signal, separating the higher harmonics from the noise simply and effectively, with the centre frequencies at the peak points.The denoising performance of the proposed method in this paper is verified by designing two categories of experiments,six groups in total,with the same vowel made from different persons and the different vowels made from the same person,where four advantages can be concluded as follow:(1)Simple calculation.(2)Good effect of extracting the main harmonics;(3)Good noise-environment transplantability; (4)Universal for different persons and different vowels;In the end in this paper,It is introduced that there are some special details expected to be improved in the proposed method,such as the extraction of the typical harmonic and small harmonic components of high frequency, average thresthold and so on.