Objective Evaluation of a Melody Extractor for North Indian Classical Vocal Performances

Obtaining accurate melodic contours from polyphonic music is essential to several music-informationretrieval (MIR) applications and is also useful from a musicological perspective. The presence of tabla and tanpura accompaniment in north Indian classical vocal performances, however, degrades the performance of common pitch detection algorithms (PDAs) that are known to provide accurate results when presented with monophonic singing voice. Recently, a melody extraction algorithm designed to be robust in the presence of the accompaniment was proposed. In this work, the same melody extraction algorithm is tested on actual professional vocal performances, facilitated by the availability of timesynchronized multi-track recordings of the voice, tabla and tanpura. The results indicate that the new method is indeed very robust to the presence of accompaniment clearly overcoming the limitations posed to melody extraction by the widely used monophonic PDAs.