Contact Information:
  Name: Yasuharu Hirasawa
  Affiliation: Graduate School of Informatics, Kyoto University
  E-mail: hirasawa@kuis.kyoto-u.ac.jp

Basic information about my algorithm:
  1. (initialization) estimate fundamental frequencies of each source
from roughly-TF-masked spectrum
  2. (main loop) use GMM sound source model and update its parameters
iteratively
  The idea of GMM sound source model is to fit sound spectra using two
kinds of GMM:
  sharp GMM fits to the harmonic components and smooth GMM fits to the
nonharmonic components.
  More details are written in my paper submitted to LVA/ICA 2012.