Computational analysis of musical style

Music is produced in an incredibly diverse number of styles across the world. These styles have many layers: for example, we can discuss the style of individual musicians (e.g. Miles Davis), subgenres (e.g. cool jazz), or genres (e.g. jazz).

Music theorists and analysts have tackled matters of style for many decades. However, there is something appealing about applying modern computational methods to these questions. In particular, computational methods can be applied at much greater scale than manual analyses, allowing us to get a more global picture of musical style and its evolution.

Recent developments in music information retrieval make it an exciting time to be computationally analysing musical style. In particular, techniques such as audio source separation (separating out a recording into its constituent instruments) and automatic music transcription (deriving MIDI from an audio file) have been experiencing great developments. As such, it is now possible to extract more musically interesting insights from audio datasets than has ever been possible before.

One approach to this project would be to go deep into a particular aspect of musical style (e.g. tonality) and study how this may be captured computationally from audio. This might involve compiling and validating techniques already in the literature, or it might involve novel signal processing and/or machine-learning research. We can then use the pipeline to study how that given musical feature varies across different kinds of musics.

An alternative approach would instead be to go deep into a given musical corpus. Here one would instead prioritise extracting a broad range of musical features, and study how these all relate to each other within and between musical styles.