Genetic learning of vocal tract area functions for articulatory synthesis of Spanish vowels

Abstract

This paper describes a multipopulation real-coded genetic approach for recovering vocal tract area functions from speech data. The kind of data analyzed is a subset of Spanish speech signals, concretely vowels from Venezuelan SpeechDat database of utterances, increasing novelty of the study. The method evolves parametric representations of speech articulators, with the goal set to minimizing acoustic distance respect to target, natural SpeechDat utterances. This distance is based on signal's formants and a measure of continuity of the area function. Subsequently, best learned functions are provided as input to an articulatory speech synthesizer, in order to generate artificial utterances, potentially and acoustically similar to the natural signals. Objective and subjective tests on these artificial signals have positively verified effectiveness of the genetic approach.