When UW faculty from computer science, engineering and statistics share knowledge, they make beautiful music together. Or, more accurately, they develop a classical music dataset that allows a computer to learn the features of classical music.
MusicNet is a collection hundreds of recordings from 34 hours of chamber music performances that can help computer algorithms deconstruct, understand, predict and reassemble components of classical music.
“Imagine I record a piece of music on a piano. I could ask my computer to take that recording and rearrange it as a string quartet,” says John Thickstun, a doctoral student in computer science and engineering who worked on the project. “This requires the computer to understand music theory and how pianos are different from string quartets. That understanding is way beyond today’s computers but we hope to start working on these problems using the dataset.”
Machine-learning researchers and music hobbyists alike can use MusicNet for note prediction, to automate music transcription, to help arrange a piece of music for different instruments, for composing, or even figuring what kind of music a person might prefer. It is the first publicly available large-scale classical music dataset with curated fine-level annotations.