STRAIGHT-Based Emotion Conversion Using Quadratic Multivariate Polynomial

Speech is the natural mode of communication and the easiest way of expressing human emotions. Emotional speech is expressed in terms of features like f0 contour, intensity, speaking rate, and voice quality. The group of these features is called prosody. Generally, prosody is modified by pitch and time scaling. Emotional speech conversion is more sensitive to prosody unlike voice conversion, where spectral conversion is the main concern. Several techniques, linear as well as nonlinear, have been used for transforming the speech. Our hypothesis is that quality of emotional speech conversion can be improved by estimating nonlinear relationship between the neutral and emotional speech feature vectors. In this research work, quadratic multivariate polynomial (QMP) has been explored for transforming neutral speech to emotional target speech. Both subjective and objective analyses were carried out to evaluate the transformed emotional speech using comparison mean opinion scores (CMOS), mean opinion scores (MOS), identification rate, root-mean-square error, and Mahalanobis distance. For Toronto emotional database, except for neutral/sad conversion, the CMOS analysis indicates that the transformed speech can partly be perceived as target emotion. Moreover, the MOS and spectrogram indicate good quality of transformed speech. For German database except for neutral/boredom conversion, the CMOS value of proposed technique has better score than gross and initial–middle–final methods but less than syllable method. However, QMP technique is simple, is easy to implement, has better quality of transformed speech, and estimates transformation function using limited number of utterances of training set.