Interactive intonation optimisation using CMA-ES and DCT parameterisation of the F0 contour for speech synthesis

Abstract

Expressive speech is one of the latest concerns of text-to-speech systems. Due to the subjectivity of expression and emotion realisation in speech, humans cannot objectively determine if one system is more expressive than the other. Most of the text-to-speech systems have a rather flat intonation and do not provide the option of changing the output speech. We therefore present an interactive intonation optimisation method based on the pitch contour parameterisation and evolution strategies. The Discrete Cosine Transform (DCT) is applied to the phrase level pitch contour. Then, the genome is encoded as a vector that contains 7 most significant DCT coefficients. Based on this initial individual, new speech samples are obtained using an interactive Covariance Matrix Adaptation Evolution Strategy (CMA-ES) algorithm. We evaluate a series of parameters involved in the process, such as the initial standard deviation, population size, the dynamic expansion of the pitch over the generations and the naturalness and expressivity of the resulted individuals. The results have been evaluated on a Romanian parametric-based speech synthesiser and provide the guidelines for the setup of an interactive optimisation system, in which the users can subjectively select the individual which best suits their expectations with minimum amount of fatigue. © 2011 Springer-Verlag Berlin Heidelberg.