Research On Emotion Recognition Of Speech Signal Based On HMM

With rapid development of Human-Computer interaction system, emotion recognition of speech has received great attention recently. Emotional factors involved in the speech signal are of great importance because they are necessary information for people's sensation of things. For example, the same sentence can convey different meaning to the same listener due to the different emotional factors it involves.In this paper, we firstly introduce the study actuality, background, research field and main application of emotion recognition of speech. Then we introduce key technologies of speech emotion recognition based on audio information and emotion classification used recently, extracting not only the features of some prosodic parameters but also with some no-prosodic parameters such as formants. During extracting the parameters of base frequency of speech, we use the algorithm of adaptive length of Hamming filter.In this paper, our classification method, the hidden Markov model (HMM) is used to classify six emotional states: happiness, anger, sadness, surprise, fear and a neutral state in which no distinct emotion is observed. The best feature vector with a dimension of eight is determined from the instantaneous features which are extracted from speech signal of our mandarin emotional speech database before being input into the HMM classifier. The instantaneous features include energy and its first and second derivative, F0 and its first and second derivative, formants and MFCC. Each hidden Markov model employed in the experiments has six states and the observation probability distribution in each state is a mixture of four normal Gaussian probability probability density distributions. Baum-Welch parameters reestimation algorithm is used to train our CHMM parameters. Segmental K-means reestimation, which can make the training results better converge on the global optimum, is conducted to train the observation probability distribution of HMM parameters. Finally we use Viterbi algorithm to recognize emotional states of the speech signal in our database, and the results from practial experiments indicate that CHMM has good data classification ability and recognition ability.The recognition rate of the six emotional states is good of which sadness and anger is better. In the end of this paper, we summarize some problems that have not been solved and the future works in this field will be discussed.