An automatic emotion recognizer using MFCCs and Hidden Markov Models

In this paper, the proficiency of continuous Hidden Markov Models to recognize emotions from speech signals has been investigated. Unlike the existing work which considers prosodic features for automatic emotion recognition, this work proposes the effectiveness of the phonetic features of speech particularly, Mel-Frequency Cepstral Coefficients which improves the accuracy with reduced feature set. The continuous speech emotional utterances used in this work have been taken from the SAVEE emotional corpus. The Hidden Markov Model Toolkit (HTK) version 3.4.1 was utilized for extraction of the acoustic features as well as generation of the models. Optimizing the acoustic and pre-processing parameters along with the number of states and transition probabilities of the Markov Models, the trials give us an average accuracy of 78% and highest accuracy of 91.25% for four emotions sadness, surprise, fear and disgust.