Part-Syllable Transformation-Based Voice Conversion with Very Limited Training Data

Voice conversion suffers from two drawbacks: requiring a large number of sentences from target speaker and concatenation error (in concatenative methods). In this research, part-syllable transformation-based voice conversion (PST-VC) method, which performs voice conversion with very limited data from a target speaker and simultaneously reduces concatenation error, is introduced. In this method, every syllable is segmented into three parts: left transition, vowel core, and right transition. Using this new language unit called part-syllable (PS), PST-VC, reduces concatenation error by transferring segmentation and concatenation from the transition points to the relatively stable points of a syllable. Since the greatest amount of information from any speaker is contained in the vowels, PST-VC method uses this information to transform the vowels into all of the language PSs. In this approach, a series of transformations are trained that can generate all of the PSs of a target speaker’s voice by receiving one vowel core as the input. Having all of the PSs, any voice of target speaker can be imitated. Therefore, PST-VC reduces the number of training sentences needed to a single-syllable word and also reduces the concatenation error.