Vocal-Accompaniment Compatibility Estimation Using Self-Supervised and Joint-Embedding Techniques

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 difficult to formulate hand-crafted rules or construct large labeled dataset perform supervised learning. Our uses self-supervised joint-embedding techniques for vocal-accompaniment compatibility. train encoders learn space where embedded feature vectors compatible pair tracks lie close those an incompatible far from other. To address lack datasets consisting pairs we generating such songs using singing voice separation techniques, which are separated into then original assumed be compatible, random not. achieved this training by constructing containing 910,803 evaluated effectiveness our ranking-based evaluation methods.