Voice emotion recognition method based on multi-fractal and information fusion

Abstract

The invention discloses a voice emotion recognition method based on multi-fractal and information fusion. The method comprises the steps that firstly, voice sample data are extracted from a voice library, and a voice sample training set and a voice sample testing set are established; secondly, a nonlinearity characteristic value used for voice emotion recognition is extracted from the voice sample training set according to the nonlinearity characteristic, wherein the nonlinearity characteristic comprises voice signal multi-fractal spectrum and a voice signal broad sense hurst index; thirdly, preprocessing is carried out on the voice sample training set, the nonlinearity characteristic value serves as the input of various weak classifiers, and all the weak classifiers are trained; fourthly, the trained weak classifiers are gathered into a powerful classifier, and the powerful classifier is tested according to voice sample signals in the voice sample testing set; fifthly, new voice signals are classified according to the tested powerful classifier, and the classifications of emotions corresponding to the voice signals are recognized. According to the voice emotion recognition method, and the accuracy of voice signal recognition is greatly improved.

Description

technical field [0001] The invention relates to a speech signal nonlinear feature extraction and a speech emotion recognition method based on the nonlinear feature, in particular to a speech emotion recognition method based on multifractal and information fusion. The present invention belongs to the field of information technology. Background technique [0002] The emotion in the speech signal is one of the important bases for judging human emotions. The speech signal contains a large amount of non-semantic information and semantic information generated according to pronunciation rules. Usually, the semantic information and non-semantic information contain language information and speech information respectively. Human emotional factors, traditional speech recognition is limited to the accuracy of speech semantics, ignoring the emotional information in speech signals, and the emotional information features of speech are usually regarded as difference noise and pattern change...

AI Technical Summary

Problems solved by technology

However, the generation of speech signals is a complex non-stationary and nonlinear process, and there is a chaotic mechanism in it. Traditional acoustic features lack the ability to describe the chaotic characteristics of speech signals.

[0005] The modeling methods of speech signals include linear modeling methods and nonlinear modeling methods, among which linear modeling methods include K nearest neighbor method, principal component analysis method, etc., nonlinear modeling methods include Hidden Markov Method (HMM), support Vector machine method (SVM), neural network method, etc. The main problem of speech recognition at present is that there are many characteristic signals that can be extracted from speech, and many methods can be used. Various signal combinations and various methods have their own advantages and disadvantages. Although Recognition can be carried out by using a variety of different signal combinations and different methods, but there are generally problems such as low recognition accuracy

Images