A Speech Transformation System for Emotion Conversion using MFCC and Modeling using GMM

Transformation of sound by using statistical techniques is significant method for a new range of digital audio effect applications. This paper talks about the data driven voice transformation algorithm, which is used to alter the timbre of a Neutral (non-emotional) voice in order to reproduce a particular emotional vocal timbre i.e. (Angry emotion). Perception based mel log spectral approximation ?lters are used to represent the speech timbre in the form of Mel Frequency Cepstral Coefficients (MFCC). The transformation function adopts a GMM (Gaussian Mixture Model) based parameterization in order to model the spectral envelope. Expectation maximization algorithm is used to find the parameters of the Gaussian mixture model. Finally a polynomial transformation function is developed for emotion transformation of the speech signal. Objective evaluation is performed by calculating Mel Cepstral distance, Prediction error and Normalized Prediction error. Mel cepstral distance between Angry and transformed Angry is found to be much lesser than mel cepstral distance between Neutral and Angry emotion. The plot of the mel cepstral distance between neutral and angry and between neutral and transformed angry also overlaps at several locations.