Is it easier to identify musicians by listening to their voices or their music? We show that for a small set of pop and rock songs, automatically-located singing segments form a more reliable basis for classification than using the entire track, suggesting that the singer's voice is more stable across different performances, compositions, and transformations due to audio engineering techniques than the instrumental background. The accuracy of a system trained to distinguish among a set of 21 artists improves by about 15% (relative to the baseline) when based on segments containing a strong vocal com-ponent, whereas the system suffers by about 35% (relative) when music-only segments are used. In another experiment on a smaller set, however, performance drops by about 35% (relative) when the training and test sets are selected from different albums, suggesting that the system is learning album-specific properties possibly related to audio production techniques, musical stylistic elements, or instrumentation, even when attention is directed toward the supposedly more stable vocal regions.
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.