Publication

Abstract : The goal of this work is to validate the impact of natural elicitation of emotions by the speakers during the development of speech emotion databases for Malayalam language. The work also proposes a Gaussian Mixture Model-Deep Belief Networks (GMM-DBN) based speech emotion recognition system. To test the effect of emotion elicitation by the speakers, two independent datasets with emotionally biased and emotionally neutral utterances are recorded in three emotions (Angry, Happy and Sad) in Malayalam language. The speech utterances of both datasets are recorded in two sessions. To develop GMM-DBN systems, the GMM models for each emotion is independently developed using Mel Frequency Cepstral Coefficients (MFCC) features and the distribution of these features for each utterance with respect to each emotion model is derived in terms of histograms with mean vectors as the frequency bins. The mean of these histograms obtained in such a manner from each emotion model is used as a feature to train the DBN. The performance of the proposed GMM-DBN system is evaluated on the developed emotionally biased and emotionally neutral datasets for Malayalam language. Based on the comparison of the emotion recognition rates obtained, a higher emotion recognition rate is observed for utterances in emotionally biased dataset which implies that the use of emotionally biased prompts during recording identify emotions more effectively. Also, the dependency of the language for the contextual prompts is observed.