A Deep Learning Exploration of Classical Music Recognition Methods

Abstract

The rapid advancement of technology has dramatically influenced our consumption of music and the arts. Readily available streaming services, such as Apple Music and Spotify, provide instant access to tens of millions of songs, while music identification tools, such as Shazam and Siri, allow users to identify unknown songs in a matter of seconds. For classical musicians, these music identification tools prove substantially more ineffective. Popular music recognition apps and programs take in "audio fingerprints" to match with an already-existing database of sounds, a database that inadequately encapsulates the classical music genre. So, while the identification method is extremely accurate, it is perhaps too accurate; errors in functionality occur when a fingerprint is not already present in the database. Classical music by nature is fundamentally structured around the interpretation of existing pieces, so in order to identify a classical piece, the database must have the specific recording of not only the piece, but also the musician. There are hundreds of interpretations of any one piece that are not present in databases. Even if the same musician were to perform the same piece twice, once for a CD, and another live, the music recognition tool would only pick up the recorded version, and only if it were in the database. Alternatively, if these tools do not render any results, the user could potentially type lyrics into a search bar as another mode of identification, but this would certainly prove difficult for a genre that is largely instrumental. Thus, the status quo presents plenty of limitations. This thesis explores machine learning and deep learning models in existing genre classification literature to ultimately identify classical pieces without the need of a “perfect match”, and instead aims for pattern recognition. The models take inspiration from Spotify’s genre classification algorithm used to create playlist recommendations. The pre-processing features are first extracted from the audio clips, and subsequently used to train various models in classifying the metadata within the classical music genre.