Harmonic-temporal-structured clustering via deterministic annealing EM algorithm for audio feature extraction
Hirokazu Kameoka, Takuya Nishimoto, Shigeki Sagayama · 2005
This paper proposes “harmonic-temporal structured clus-tering (HTC) method”, that allows simultaneous estima-tion of pitch, intensity, onset, duration, etc., of each under-lying source in multi-stream audio signal, which we ex-pect to be an effective feature extraction for MIR systems. STC decomposes the energy patterns diffused in time-frequency space, i.e., a time series of power spectrum, into distinct clusters such that each of them is originated from a single sound stream. It becomes clear that the problem is equivalent to geometrically approximating the observed time series of power spectrum by superimposed harmonic-temporal structured models (HTMs), whose parameters are directly associated with the specific acoustic charac-teristics. The update equations in DA(Deterministic An-nealing)EM algorithm for the optimal parameter conver-gence are derived by formulating the model with Gaussian kernel representation. The experiment showed promising results, and verified the potential of the proposed method.