Music Accompaniment Generation with Instrumental Solos for Pop Music

When generating instrumental accompaniments for a popular song’s vocal melody, the accompaniment needs to be able to support the lead melody sung by the vocalist so that the lead melody can sound more colorful with the accompanying instrumental melody. In a popular song, the lead melody or the singer may have”break moments” where the lead melody or singer is silent. In these moments, the instrumental accompaniments generated by state-of-the-art models do not deviate from what they normally generate, which can cause listeners to lose interest. Thus, because there is no lead melody, there is nothing to accompany. We propose an interpolation composition method based on the Transformer model that can improve the current accompaniment generating models by adding an instrumental solo generation module whenever the singer is absent. Overall, our Interpolation Transformer was able to generate instrumental solos that can combine with the original accompaniment to improve upon the baseline accompaniment generation model. We also learn that including an instrumental solo in the parts where the singer is quiet indeed increases the likelihood that the accompaniment is preferred twice as much by human evaluators compared to an accompaniment without an instrumental solo.