A fundamental study of structure-to-speech conversion

Speech acoustics vary due to differences in age, gender, vocal tract length, microphone, and so on. The authors recently have proposed a structural and abstract representation of speech, where these variations were effectively removed. In this study, a framework of speech synthesis based on this structural representation of speech is proposed. In the proposed framework, a system needs a “speech gestalt” of one utterance, properties of vocal tract length of speaker and properties of transmission of microphone. Using these information, acoustic signals to which hearers listen are generated. This framework can be regarded as a model of vocal imitation of infants. For a fundamental consideration of this framework, the authors considered this framework as a problem of searching cepstrum space for the solutions under some constraints in this report. As results of experiments, speech samples which have proper phonological characteristics were synthesized.