Singing voice conversion method based on many-to-many eigenvoice conversion and training data generation using a singing-to-singing synthesis system

The voice quality (identity) of singing voices is usually fixed in each singer. To overcome this limitation and enable singers to freely change their voice quality using signal-processing technologies, we propose a singing voice conversion method based on many-to-many eigenvoice conversion (EVC) that can convert the voice quality of an arbitrary source singer into that of another arbitrary target singer. Previous EVC-based methods required parallel data consisting of song pairs of a single reference singer and many prestored target singers for training a voice conversion model, but it was difficult to record such data. Our proposed method therefore uses a singing-to-singing synthesis system called VocaListener to generate parallel data by imitating singing voices of many prestored target singers with the system's singing voices. Experimental results show that our method succeeded in enabling people to sing a song with the voice quality of a different target singer even if only an extremely small amount of the target singing voice is available.