Eigenrhythms: Drum pattern basis sets for classification and generation

We took a collection of 100 drum beats from popular music tracks and estimated the measure length and downbeat position of each one. Using these values, we normalized each pattern to form an ensemble of aligned drum patterns. Principal Component Analysis on this data set results in a set of basis ‘patterns’ that can be combined to give approximations and interpolations of all the examples. We use this low-dimension representation of the drum patterns as a space for classification and visualization, and discuss its application to generating continua of rhythms. Our classification results were very modest – about 20% correct on a 10-way genre classification task – but we show that the projection into principal component space reveals aspects of the rhythm that are largely orthogonal to genre but are still perceptually relevant.

• Type: Conference
• Year: 2004
• Where: ISMIR
• Authors: Dan Ellis, John Arroyo
• Keywords: Aligned Drum Patterns, Drum Patterns, ISMIR 2004, Principal Component