CONTENT-BASED TRANSFORMATION OF THE EXPRESSIVITY IN SPEECH
Xavier Rodet · 2007
In this paper we describe a transformation system for speech expressivity. It aims at modifying the expressivity of a spoken or synthesized neutral utterance. The phonetic transcription, the stress level and the other information about the corresponding text supply a sequence of contexts. Every context corresponds to a set of parameters of acoustic transformation. These parameters change along the sentence and are used by a phase vocoder technology to transform the speech signal. The relation between the transformation parameters and the contexts is initialized by a set of rules. A Bayesian network transforms gradually this rule-based model into a data-driven model according to a learning phase involving an expressive French speech database. The system functions for French utterances and several acted emotions. It is employed at artistic ends for multi-media applications, the theater and the cinema.