Extraction Of Speech Features Based On Human Auditory Characteristics

Which this article formulates is divided into 3 parts: hearing masking, equal loudness and property of cochlear frequency-division. The main contributions as below:1) An approach carving out bark frequency band with mel frequency mark is brought forward and acquired a barkmel sheet. In research on the basis of hearing masking, a concept of masking and speech recognition method by traditionally using masking feature. Traditional method also includes removing stable noise signal from the speech by means of spectral subtraction. This article formulates the relation between mel Frequency and bark Frequency band and acquired a barkmel sheet. In experiment tothis sheet, cepstrum exaltation is also done to MFCC. First and second difference will be done to MFCC and one energy frame so that acquire 39 dimension feature parameters to optimize masking feature in picking up MFCC.2) United weighting of SNR and equal loudness property is brought forward and a new weighting filter is designed according to the new ISO criterion. In research based on equal loudness property of human auditory system, we discovered that common equal loudness weighting divided speech signal into sections by spectrum and each of sections is weighted respectively, If the SNR of one speech section is low, no doubt, by this method, noise will impact more on signal and make against enhancing recognition accuracy rate. This article formulates that after signal passed equal loudness filter, weighting should be done to the weighted Frequency bands by using SNR . Besides, in this article, a new weighting filter is worked out according to the new ISO equal loudness contour.3) Adding equal loudness weighting into subband is brought forward. In research on subband technology on the basis of property of cochleafrequency-division, this article analyzed the theoretic evidence that subband frequency-division technology is better than common method under the condition of noise followed by bringing out the method that adding equal loudness property into sub-band to enhance recognition accuracy rate. In this article, SNR weighting> method of simultaneous weighting for HMM model and characteristic coefficient to mate feature space with model space better, which have been mentioned in other references, are also researched. Particularly in introduction on SNR weighting approach, the acquired the value of SNR can be more accurate by utilizing SNR smooth technology.