Speech Analysis for Identification of Emotion

Emotions play a significant role in expressive speech. The emotional state of a speaker is demonstrated by physiological signs in the form of changes in respiration, phonation and articulation. This research project examines whether it is possible extract features of speech to identify distinct aspects of speech samples to represent varying emotion in speech by examining the prosodic features of each emotion. This research project follows techniques used in speech analysis. Speech analysis has typically used specific algorithms to detect features in samples to link a speaker to their sample, either for identification of the speaker, or identification of the words being spoken (Juslin, and Scherer, 2008). In particular, parameterisation of samples using Linear Predictive Coding (LPC) and Mel Frequency Cepstral Coefficients (MFCC) have been successfully applied in speech analysis (Hess 1983). These parameterisations are used to identify and distinguish features, which can be used to classify samples according to speaker. This research project continues with this theory, applying the LPC and MFCC techniques to determine if speech samples can be classified according to one of eight emotions. This outcome of this project demonstrates that parameters used in speech recognition can be applied to classifiy speech samples into specific emotions. By applying the LPC parameter to the speech samples, general features of each emotion were able to be extracted, and the resulting features were compared and classified and the research project achieved 100% correct classification for some emotional states and overall an accuracy of 72% across the eight emotions