Transformation of short-term spectral envelope of speech signal using multivariate polynomial modeling

Speaker transformation uses a mapping between the signal parameters derived from a set of phrases spoken by two speakers to modify the speech signal of one speaker to make it perceptually similar to that of the other speaker. In spectral warping and interpolation technique, the transformation function can be estimated using lesser data, but it needs a different transformation function for each acoustic class. Earlier reported investigations have used class dependent multivariate linear regression or univariate cubic interpolation. In this paper, a technique for modifying spectral characteristics, using a multivariate polynomial modeling for source-target mapping of the spectral parameters, is presented. Harmonic plus Noise Model (HNM) has been used for analysis-synthesis because of ease of the time and frequency scaling.