Adversarial Unsupervised Domain Adaptation for Harmonic-Percussive Source Separation

This letter addresses the problem of domain adaptation for task music source separation. Using datasets from two different domains, we compare performance a deep learning-based harmonic-percussive separation model under training scenarios, including supervised joint using data both domains and pre-training in one with fine-tuning another. We propose an adversarial unsupervised approach suitable case where no labelled (ground-truth signals) target is available. By leveraging unlabelled (only mixtures) this domain, experiments show that our framework can improve on new without losing any considerable original domain. The also introduces Tap & Fiddle dataset, dataset containing recordings Scandinavian fiddle tunes along isolated tracks “foot-tapping” “violin”.