An extended audio-fingerprint method with capabilities for similar music detection

Content-based Audio Identification consists of retrieving the meta-data (i.e. title, artist, album) associated with an unknown audio excerpt. Audio fingerprint techniques are amongst the most efficient for this goal: following the extraction of a fingerprint from the unknown signal, the closest fingerprint in a reference database is sought in order to perform the identification. While being able to manage large scale databases, the recent developments in fingerprint methods have mostly focused on the improvement of robustness to post-processing distortions (equalization, amplitude compression, pitch-shifting,...). In this work, we describe a novel fingerprint model that is robust not only to the classical set of distortions handled by most methods but also to the variations that occur when a title is re-recorded (live vs studio version in particular). As a result our fingerprint method is able to identify any signal that is an excerpt of one of the references from the database or that is similar to one of the references. The issue that we cover thus lies at the intersection of audio fingerprint and cover song detection, meaning that the functional perimeter of our method is substantially larger than the classical audio fingerprint approaches.