Multipopulation genetic learning of midsagittal articulatory models for speech synthesis

摘要

This paper discusses an application of multipopulation continuous genetic algorithms to learning of vocal tract configurations on a midsagittal plane. Speaker dependent and independent target signal corpora are formed and processed by the genetic approach, which evolves populations of articulatory vectors in order to approximate the acoustic traits of artificial utterances to those of natural signals in the corpora. Analyzed signals correspond to venezuelan spanish speakers, increasing novelty of the study. Subjective evaluations have confirmed effectiveness of the method, reaching a 19% recognition error for speaker independent trials, and no error for the speaker dependent case.