On the perceptual relevance of objective source separation measures for singing voice separation

Singing Voice Separation (SVS) is a task which uses audio source separation methods to isolate the vocal component from the background accompaniment for a song mix. This paper discusses the methods of evaluating SVS algorithms, and determines how the current state of the art measures correlate to human perception. A modified ITU-R BS.1543 MUSHRA test is used to get the human perceptual ratings for the outputs of various SVS algorithms, which are correlated with widely used objective measures for source separation quality. The results show that while the objective measures provide a moderate correlation with perceived intelligibility and isolation, they may not adequately assess the overall perceptual quality.