You’d think that jazz, with its emphasis on live improvisation and its loose structure, would be the hardest genre of music for AI to analyse, wouldn’t you? Not so, according to a new research paper published in Nature Machine Intelligence.
The research, which was conducted at the University of Cambridge by a team led by one Huw Cheston, has shown that large language models can identify jazz pianists and even reveal the musical ‘fingerprints’ that made them stand out.
Cheston’s team fed the machine 84 hours of recordings from 1,629 different performances by 20 pianists. Among those that chosen for the study were Bill Evans, Oscar Peterson, Thelonious Monk, Chick Corea, Keith Jarrett, McCoy Tyner and Ahmad Jamal. The recordings were then converted into MIDI and displayed on a piano roll, which showed digitally what notes were played, when and at what pitch.
And wouldn’t you know it, the bots had a pretty good success rate. The best-performing model identified the performers with a rate of 94.4% accuracy. Interestingly, when tested individually, it was the composition’s harmony that gave the most accurate predictions, followed by rhythm and melody, while dynamics was the least accurate.
All very clever, of course. But what use is all this in the real world, you might well ask? Well the authors of the research say it may well have a number of uses. AI used in this way could suggest “possible authors for unattributed works or distinguishing authentic pieces from forgeries. It could also “highlight the networks of influence between artists, as well as reveal insights into the creative process itself.”
“Finally,” they say, “it can also be useful in educational contexts, helping to train artists to understand the distinctive voices of important figures within their particular discipline.”
If you’re interested in dipping further into the study head over to Nature Machine Intelligence’s website. There's also a nice little web app that enables you to explore the playing style of each pianist.
