Speech Emotion Recognition Based On A Hybrid Of HMM And RBF

The rapid developme nt of comp uter has a huge impact on huma n society, thecommunication between huma n and co mputer become the key techno logy ofinte llige nt and huma nized computer system, the new type of human-co mputerinteraction technolo gy is gradually becoming a research hotspot. Stud ies show that theability of "e motiona l inte lligence " is the one of the important proble ms needed to besolved in the huma n-computer interaction. Currently, the research about emotiona linformatio n processing is deep-going, in especia lly, the studies of e motio ninformatio n processing in speech s igna l receive more attention from researchers.Speech emotion recognition is an important branch of the "affective computing", andit is developing rapidly with a broad prospects.At first, this paper has executed pretreatment s uch as pre-emphas is and enfra me.Statistica l ana lyzed about energy, pitch a nd duration at d ifferent e motiona l states.Amo ng the extracted features, selected and norma lized energy, pitc h freque ncy, linearprediction coeffic ie nt (LPC), linear prediction cepstrum coeffic ie nt (LPCC), and Melprediction cepstrum coeffic ient (MPCC) fro m emotiona l speech signa ls as the input ofthe identification phase.Stud ied at the two identification methods inc lude Hidden Markov model (HMM)and radia l basis function neura l network (RBF), and ana lyzed the ir advantages anddisadvantages. In this paper, in order to achie ve the purpose of comp le mentaryadvanta ges, we design a hybrid model class ifier for e motio n speech recognition basedon mode ling sequences by HMMs, and making dec is ion by RBF. Each HMM modelone emotio n. HMMs, training the emotion utterances, adopted Baum-Welc hre-estimatio n algorithm based on Maximum Likelihood probabilities training criterio n.In the process of training, Viterb i algorithm, concentrated on the best path through themodel, eva luated the like lihood of the best match between the give n HMMs and thegive n speech observatio ns, then obtaining the optima l state sequence. The n, this paperselected Legendre po lyno mials to be the orthogona l bases function to get the equa ldimens ion sequence. ANN had been emp loyed to make a decis ion. The recognitio nresult o f the hybrid class ification has been compared with the iso lated HMMs and theiso lated RBF by CASIA mandarin speech emotion corpora. In the speaker-dependentand speaker-independent, the average recognition rates have reached80.45%and 69.31%respective ly. Fina lly, used majority voting fus ion for HMM, RBF and hybridmodel improved the recognition.