Radio, Radio is an ongoing project of data collection and 'learning by observation' for music and language, working towards "music acquistion:" modelling the evolution of music and individual learning of music parameters. We have set up a big cluster of multiprocessor machines that share the job of listening to as much Internet radio as they can handle. At certain times of the day, some of the machines begin analysis on the collected data using machine learning and feature extraction techniques to pull out different levels of musical structure. These musical models get better every day, and can be used in such music retrieval and understanding tasks as recommendation, segmentation, classification, and synthesis.

At the same time, our Klepmit "community metadata" (paper) crawlers read all about the artists they have just heard using language processing techniques, building up a cultural knowledgebase of the current music. The link between the words and the music is usually where we find the best stuff-- it's the community view on audio, not just the audio itself that defines music's meaning. (paper)

The same system does night duty as DJ for Eigenradio.

Colophon: We're currently using PLAPACK over BLACS over ATLAS on Linux. We're also big fans of PLab. Super-UROP David Alexander did a lot of the hard work getting our parallel kernel Gram matrix and inversion routines going.

Brian Whitman