We present a live drum separation system for a specific target drumset to be used as a front end in a
complete live drum understanding system. Our system decomposes drum note onsets onto spectral drum
templates by adapting techniques from non-negative matrix factorization. Multiple templates per drum are
computed using a new Gamma mixture model clustering procedure to account for the variety of sounds that
can be produced by a single drum. This clustering procedure imposes an Itakura-Saito distance metric on
the cluster space. In addition, we utilize “tail” templates for each drum which greatly improve the separation
accuracy when cymbals with long decay times are present.