The Research On Feature Extraction And Recognition Of Emotional Speech

With the development of human-computer interaction technology, more and moreresearchers pay attention to the emotion in the speech signal. Emotional feature extraction isone of the important branches of research in speech signal，the present method in emotionalfeature of the speech can not reflect the emotional characteristics effectively and has a lowrecognition rate , aiming to this problem, this paper has done research works as bellow:1. According to the characteristics of speech, this paper extracts some usual speechfeature, such as fundamental frequency, formant, Mel Frequency Cepstral Coefficient. Thenanalyses these emotion features with experiments. Experimental results show that thesespeech features have emotion characteristics indeed.2. The recognition rate, which we get from distinguishing different emotion speechonly depend on speech features, is very low. To solve this problem, this paper gets theintrinsic mode function of speech emotional signals by the Empirical Mode Decomposition,then extracts the instantaneous frequency and instantaneous amplitude of IMF using Hilberttransform. Through experimental analysis, we find that the same order IMF of the differentpeople in different emotional states is not same, even the same one IMF. Making all of theIMF as the emotion features of speech emotion recognition do not get the best result,meanwhile, recognition performance will get best when we use the first four IMF features. So,this paper just extracts the first four IMF features, use their instantaneous frequency andinstantaneous amplitude as new emotion feature.3. In this paper, we use a way of support vector machine to identify emotion speech.First, set happiness anger as a class, sadness and calm as a class, training and recognizingemotion speech, then train and recognize each part separately. Experimental results show thatthe method has a good recognition effect, compared with other methods, the recognition rateis improved effectively.