Metric Learning with Sequence-to-sequence Autoencoder for Content-based Music Identification

Abstract

Content-based music identification is an active research field that involves recognizing the identity of a musical performance embedded within an audio query. This process holds significant relevance in practical applications, such as radio broadcast monitoring for detecting copyright infringement. Various approaches for content-based music identification have been explored in the existing literature, yielding diverse levels of performance. However, despite the considerable attention dedicated to this area, no attempts have been made to leverage the dynamical nature of musical works coupled with the modern advances in machine learning such as metric learning for content-based music identification. In this paper, we propose a novel approach that encodes the dynamic nature of musical performances into the latent space of a sequence-to-sequence auto-encoder network. The learning objective is further enforced with the metric learning for music similarity measurement. The proposed model is extensively evaluated by testing it with 14 distortions of the same musical performance. The experimental results demonstrate a substantial increase of 31.71% in hit-rate over the baseline established using related work found in the literature. These findings highlight the potential of our approach to significantly improve content-based music identification, thereby offering promising applications in various practical scenarios.