Research On Speech Emotion Recognition Technology

Along with the rapid development of information and communication technology and mobile Internet technology, people are eager to communication with computer in an intelligent, emotional, humanization way. Speech is the most direct means of communication, and it is also the main carrier of human emotions. Now research on speech emotion recognition technology is a new research direction, which not only has a vital significance to the human computer interaction, but also has an important influence on the artificial intelligence.Based on some existing methods of speech emotion recognition technology which have been studied and analysized, this thesis studies the emotional features from three angles, namely supra-segmental feature, spectral feature and a feature based on multi-resolution analysis of critical band; Lastly, the first attempt to adopt projection Dictionary Pair Learning(DPL) algorithm to solve the problem of speech emotion recognition. Researching work of this thesis mainly includes the following several aspects:1. Emotional features are studied in this thesis, including the two following aspects: 1) Supra-segmental feature and spectral feature are extracted. Supra-segmental feature includes fundmental frequency(F0) and loudness; spectral feature includes MFBECS and LSF. The method of F0 extraction is studied, in view of the problem of frequency errors existing in SHS algorithm, an improved SHS algorithm is proposed. 2) Because not all parts of the spectrum can affect human perception system, a feature based on multi-resolution analysis of critical band, GPWP feature, is introduced. The wavelet packet basis function used in the GPWP is studied, and the result shows that the recognition performance of coif2 basis function is the best.2.This thesis studies speech emotion recognition based on Sparse Representation Classification(SRC). But SRC exists the problems of slow speed and unsatisfactory recognition effect in solving the speech emotion recognition, the first attempt to adopt DPL algorithm to solve the problem of speech emotion recognition.3. In this thesis, three speech emotion databases, Emodb, Polish and eNTERFACE’05, are used to carry out the experiment. Firstly, the recognition performance of GPWP feature is studied, and the result shows that the recognition effect of GPWP is better than the other four features; Secondly, the combination of emotional features of the thesis and related literature are compared, the results show that the combination of emotional features of this thesis is better; Lastly, the number of atoms of DPL is studied, from two aspects of time performance and recognition performance, the method is compared with the other four recognition methods, namely SRC, SVM, JSLRR and CRC, the result shows that the DPL method not only has a better recognition performance, but also has a better time performance.