Singing Expression Transfer

We present a vocal processing algorithm to automatically transfer singing expressions from one voice to another for a given song. Depending on singers' competence, a song can be rendered with great variations in terms of local tempo, pitch and dynamics. The proposed method temporally aligns a pair of singing voices using melodic and lyrical features that they have in common. Then, it conducts time-scale modification on the source voice according to the time-stretching ratio from the alignment result after smoothing. Once the two voices are aligned, the method modifies pitch and energy expressions of the source voice in a frame-by-frame manner using a pitch-synchronous overlap-add algorithm and a simple amplitude envelope matching. We designed our experiment to transfer singing expressions from a highly technical singer to a plain singer. The results show that our proposed method improves the singing quality effectively.

The audio player used for this demo is modified version of waveform-playlist.