The Research On Speech Emotion Recognition Based On Multi-Weights Nerual Network

Dealing with the speaker's emotion is one of the latest challenges inspeech technologies. Various recent studies have dealt withcharacterization of emotional speech and automatic classification ofemotional utterances. It should be noted that the expected recognition rateis much lower than what is expected from speech recognition tasks,though a variety of methods have been proposed. In this paper, a novelspeech emotion recognition, that is speech emotion recognition based onmulti-weights neural network (MWNN), has been proposed.Firstly, the graphic geometry theory for feature space sample pointsand its properties have been presented in this paper, while the featurespace sample point's properties have been researched through theconcepts of graph theory. Then the algorithm based on convex hull forsub-graph is presented. Finally the multi-weights neuron based on graphic geometry theory is proposed with distance among sub-graph's samplepoints as the weight. We discussed the characteristics of emotional speech,and investigated the methods for feature extraction and data compression.After referring various emotional speech theories and considering effectson speech emotion recognition from each emotion feature parameters, weconfirmed the most effective parameters as the features for therecognition task and propose algorithms of feature extraction after speechpretreatment. According to the parameters, we considered the algorithmof speech emotion recognition based on multi-weights neurons fortraining and recognition.The experiment results show that the difference of recognition ratebetween MWNN and SVM model will decrease as the number oftraining samples increases. Both of their rates tend to be a peak value ifthe number of training samples is infinite. But the recognition rate ofMWNN-based method is always higher than that of SVM-based methodwhen the neurons are enough because the fore method can describe themorphological distribution of emotional speech samples in the highdimensional space while the latter one just partitions the highdimensional space into a few areas. Especially when the number oftraining samples is very few, the precision of SVM-based method isworse than that of MWNN method due to the limitation of supportvectors.