Prosody Transfer from a Small Amount of Voice using Fine Tuning

Text-to-Speech(TTS) is a technology that generates corresponding speech from textual input. A large amount of audio data is needed to build a TTS model generating natural voices. Therefore, if the amount of available data of the target is small, you need to take measures to reduce the amount of data required for training. The purpose of this study is the prosodic transfer of the target speech with a small amount of available data by fine-tuning the base model between the same languages. We investigate how much prosodic information was transferred. As prosodic information, we focus on the speaker’s voice quality, and dialect. We evaluate dialect from "pronunciation" in the English TTS model and "intonation" in the Japanese TTS model. At present, it was possible to transfer the voice quality of same-sex speakers from a small amount of voices for about 15 minutes. Regarding the dialect, the difference in pronunciation could not be transferred, but the intonation could be transferred.