Optimal Phonemic Environmental Range to Select VCV Instances for VCV Speech Synthesis by Rule

Abstract

We proposed two selection methods, 1) selection method by using phonemic environmental resemblance score (PER method), and 2) selection method by searching minimal connective distortion path (MLD method), for small scale speech synthesis system. PER method requires phonemic environmental information for each VCV instance in a VCV unit dictionary. This paper investigated experimentally to what extent we can reduce the phonemic environmental information with keeping high quality of synthesized speech. We verified that two phonemes frontward and one phoneme rearward range to a current VCV instance is enough to synthesize similar quality of speech as five phonemes frontward and five phonemes rearward. This result gives an experimental basis on minimizing a size of VCV unit dictionary.