Song Identification From Live Music Using Siamese Convolutional Neural Networks

Abstract

The task of live music song identification is about identifying the underlying song being performed in a live recording track. Automated song identification systems have long been used in applications like Shazam and Youtube content-id as tools for music information retrieval and copyright infringement detection, respectively. However, the fingerprinting-based methods used in audio copy detection perform rather poorly when musical changes are concerned. Two recordings presenting the same song can have differences in their tempo, song key, instrumentation, arrangement, chords, melodies, rhythms and lyrics. Live music environment brings additional complexity to the task as crowd noises, improvisation of performing artists and inaccuracies in timing or pitch of played notes can further alter the song. Recent studies about cover song identification have managed to develop deep learning methods that are robust to musical changes. However, it is unknown how well these methods work with live music due to the lack of ongoing research about the topic.