Singing pitch extraction and voice separation from music accompaniment using trend estimation and tandem algorithm

The interdependency between pitch estimation and pitch-based separation creates a problem for pitch estimation and voice separation. It is observed that the target pitch can be estimated from a few harmonics of the target signal. The trend estimation algorithm is proposed to bound the singing pitch in a plausible range and an extended tandem algorithm is also proposed that estimates target pitch and separates singing voice from music accompaniment iteratively. By coupling with the proposed trend estimation, the tandem algorithm is improved significantly for both pitch estimation and voice separation in musical recordings. Systematic evaluation shows that the proposed algorithm performs significantly better than a previous CASA and a model-based system.