We propose a learning-based method of estimating the compatibility between vocal and accompaniment audio tracks, i.e., how well they go with each other when played simultaneously. This task is challenging because it is difficult to formulate hand-crafted rules or construct a large labeled dataset to perform supervised learning. Our method uses self-supervised and joint-embedding techniques for estimating vocal-accompaniment compatibility. We train vocal and accompaniment encoders to learn a joint-embedding space of vocal and accompaniment tracks, where the embedded feature vectors of a compatible pair of vocal and accompaniment tracks lie close to each other and those of an incompatible pair lie far from each other. To address the lack of large labeled datasets consisting of compatible and incompatible pairs of vocal and accompaniment tracks, we propose generating such a dataset from songs using singing voice separation techniques, with which songs are separated into pairs of vocal and accompaniment tracks, and then original pairs are assumed to be compatible, and other random pairs are not. We achieved this training by constructing a large dataset containing 910,803 songs and evaluated the effectiveness of our method using ranking-based evaluation methods.
ASJC Scopus subject areas
- コンピュータ サイエンス（全般）