Study Of Front-End Process In Speech Feature Extraction And Optimized Frame Algorithm In Noisy Environment

Feature extraction is a key technology for speech recognition. Inrecent years, the feature extraction based on human auditory model hasbeen attracting more and more attention. This is because the recognitionability of the human ear is very well, even in the noisy environment.Among many kinds of feature extraction methods, Zero-crossing with PeakAmplitude (ZCPA) feature extraction is just based on the auditory model ofhuman ear. This model uses zero-crossing interval to represent signalfrequency information and amplitude to represent intensity information.Then frequency information and amplitude information are combined toform the complete output of speech signal. This paper aims to present somekinds of the improved ZCPA feature on the basis of above-mentioned.This paper first introduced wavelet transform as analyzing tool anddiscussed the theory of wavelet transform and its characteristics in timedomain and frequency domain. A method to design a combined waveletsfilter was proposed by taking frequency shift and single waveletsuperposition. Using combined wavelets, the low-pass, high-pass,band-pass filters can be easily designed by selecting the proper parameters of the wavelet. The experiments showed that this filter had advantages ofsimple algorithm, good frequency characteristics and was easily forcomputer software.This paper used wavelet theory in ZCPA feature extraction front-endprocess and introduced new feature extraction methods used Gauss waveletfilter and combined wavelets instead of the FIR filters in original ZCPAmethod respectively. According to critical frequency band of human earsthe Gauss wavelet filter and combined wavelets filters are desigried bystudying human auditory characteristic. The method of choosing scaleparameter in designing Gauss wavelet filter and the different characteristicsof combined Wavelets filters with different number of single wavelet havebeen given a detailed discussion. The RBF neural net is used in back-endtraining and recognition course. The results showed that new feature hadhigher recognition rate and better robustness than traditional feature.Then this paper analyzed and improved the frame processing of ZCPAfeature extraction. In the processing of speech signal feature extraction,speech signal is usually operated by framing. Frame, especially the lengthof frame, can affect the final recognition rate directly. As a result, theoptimized frame is proposed. This paper detailedly discussed the influencesthe different length of frame makes on recognition rate on the basis ofZCPA feature extraction, simulating the improved system. Since a lot ofhigh frequency information are missing in the process of obtaininginformation through computing the upward-going zero-crossing rate inoriginal ZCPA method and the extracted information more conform with the hearing characteristics of human, this paper got the difference of speechsignal to obtain the high frequency information by computingupward-going zero-crossing rate of the difference signal. At the same time,optimized frame can compensate some high frequency information. Theweighting matrix is used to weight the density information for the purposeof making it more correspondent with the human hearing characteristics. Inthese two ways, it improves the deficiency of ZCPA. Experimentseventually show that the improved algorithm had higher recognition rate.