SpeakBySinging: Converting singing voices to speaking voices while retaining voice timbre

This paper describes a singing-to-speaking synthesis system called "SpeakBySinging" that can synthesize a speaking voice from an input singing voice and the song lyrics. The system controls three acoustic features that determine the difference between speaking and singing voices: the fundamental frequency (F0), phoneme duration, and power (volume). By changing these features of a singing voice, the system synthesizes a speaking voice while retaining the timbre of the singing voice. The system first analyzes the singing voice to extract the F0 contour, the duration of each phoneme of the lyrics, and the power. These features are then converted to target values that are obtained by feeding the lyrics into a traditional text-to-speech (TTS) system. The system finally generates a speaking voice that preserves the timbre of the singing voice but has speech-like features. Experimental results show that SpeakBySinging can convert singing voices into speaking voices whose timbre is almost the same as the original singing voices.