Audio Sources Separation by Clustering Techniques

[Background] Audio sources, including human vocal cords and musical instruments usually are based on a harmonic oscillator that produces vibrations on a series of integer multiples of a base frequency. The base frequency defines the pitch of the sound produced. Due to different characteristics of these sources, even if we have 2 sounds from 2 different sources whose base frequencies are the same, their power distributions over these harmonics are different. We human ears can identify the differences in the distributions and tell that the instruments produce different timbres. To us, timbres are signatures of sound sources. However, these signatures are not that straightforward to computers. Since these signatures are described in the spectrum domain, my starting point to tackle this problem is from the spectrum of audio signals.