Transformation of LF parameters for speech synthesis of emotion: regression trees

Abstract

This paper outlines an approach to modelling the dynamics of voice source parameters as observed in the analysis of emotional portrayals, by a male speaker of Hiberno-English. The emotions portrayed were happy, angry, sad, bored, and surprised, as well as neutral. The voice source parameters extracted from emotionally coloured repetitions of a short utterance ? by means of inverse filtering followed by source model matching ? were modelled using classification and regression trees. Regression trees were built using the voice source parameters of the neutral repetition of the same short utterance, in order to transform the voice source parameters from neutral to one of the five emotions. Re-synthesis of emotion-portraying utterances using transformed voice source parameter dynamics resulted in synthesised utterances which were confirmed by listening tests to represent the targeted emotion categories. The results suggest that the addition of dynamic voice source information in parametric synthesis of emotion will improve the quality of emotion synthesis.