A Study Of Speech Feature Extraction Based On Manifold Learning

The goal of speech recognition is to let the computer can understand what the human is speaking,it is an important research direction of the speech signal processing.Speech signal is the main object that speech recognition is to study With the development of computer technology,multimedia technology,digital signal processing technology and the development of Internet technology,peo-ple have high expectations for the development of speech recognition technology.Speech feature extraction is one of the key technologies of speech recognition.In this paper,a new feature extraction method based on manifold learning is pro-posed by systematically studying the principle of speech generation and auditory mechanism.The manifold algorithm is a non-linear dimensionality reduction method developed in recent years.At present,most of the related research of manifold learning has achieved good results in image processing,face recognition and handwriting recognition,but very few studies on acoustic signal processing,specially on speech recognition field.The characteristic parameters MFCC and LPCC of speech recognition are based on the theory of linear system,but hu-man's speech system is not a linear time invariant system,and MFCC and other characteristics are difficult to reflect the intrinsic features of the speech signal.The purpose of manifold learning is to explore the geometric structure of data,or to find the intrinsic features of nonlinear data.The use of manifold learning method to study the speech signal is to find out the inherent characteristics of the speech signal,and to find the low-dimensional manifold of the speech data,and improve the accuracy of ASR system.By improving the accuracy of ASR system,and then improve the efficiency of artificial intelligence,language input,identification and other applications,has a certain practical significance.In this paper,we first introduce the theory of manifold learning,the principle of speech generation and the common cue-based speech feature parameters.In this paper the existence of a low-dimensional curved manifold,which structure to voiced speech sounds,is analyzed with a simple model for the vocal tract proven useful in the tradition of acoustic modeling of voiced phoneme production.Finally,a method of speech feature extraction based on manifold learning is proposed.The innovation of this method is to link the manifold learning technology with the vocal principle and the auditory mechanism of the human.The phonetic features extracted by this method have better performance than traditional feature extrac-tion methods in the separability of phonetic phonemes and phoneme clustering.This method provides a new choice for the extraction of feature parameters in speech recognition applications such as artificial intelligence,language input and identity recognition,which provides reference for relevant researchers.