News release

From: Springer Nature

Artificial intelligence: Revealing the fingerprints of jazz musicians

Machine learning models can identify jazz pianists from recordings and reveal the musical ‘fingerprints’ that make individual performers recognisable, according to a study investigating 20 famous jazz pianists. The research, published in Nature Machine Intelligence, could offer insight into artist attribution and style, cultural heritage, and music education.

Musicians can often be recognised by distinctive patterns, or ‘fingerprints’, in their work, which can include harmonic progressions, rhythmic structures, and melodic motifs. Jazz offers an opportunity to study these fingerprints because it brings together composition, improvisation, and performance that can be especially unique to a specific performer. However, computational analysis of jazz has been difficult because most performances exist as audio recordings, and models trained directly on audio can be hard for humans to interpret due to variables such as the environment of the recording and the types of equipment used to record.

Huw Cheston and colleagues trained supervised-learning models to identify performers from a curated dataset of 84 hours of recordings from 1,629 performances by 20 famous jazz pianists. These included recordings of Bill Evans, Oscar Peterson, Thelonious Monk, Chick Corea, Keith Jarrett, McCoy Tyner, and Ahmad Jamal. The recordings were converted into MIDI ‘piano roll’ format, which is a digital representation that shows when notes are played and at what pitch. The best-performing model identified performers with 94.4% accuracy, while a more interpretable model that separated melody, harmony, rhythm, and dynamics reached 91.3% accuracy. When tested individually, harmony gave the most accurate predictions, followed by rhythm and melody, while dynamics was the least accurate.

The authors suggest that their approach could help researchers explore the musical patterns that make performers distinct, including features that align with existing scholarly literature and others that have not previously been discussed. They have also released open-source models and a web application to explore the results. However, they note that musical dimensions such as melody and harmony can overlap, and that MIDI piano rolls cannot capture all aspects of jazz performance, including tonal and timbral qualities such as vibrato and pitch-bending. Future work could extend the approach to other genres, instruments, historical contexts, and less well-known or historically under-represented musicians.