Atsuo Hiroe

Solution of permutation problem in frequency domain ICA, using multivariate probability density functions

Conventional Independent Component Analysis (ICA) in frequency domain inherently causes the permutation problem. To solve the problem fundamentally, we propose a new framework for separation of the whole spectrograms instead of the conventional binwise separation. Under our framework, a measure of independence is calculated from the whole spectrograms, not individual frequency bins. For the calculation, we introduce some multivariate probability density functions (PDFs) which take a spectrum as arguments. To seek the unmixing matrix that makes spectrograms independent, we demonstrate a gradient-based algorithm using multivariate activation functions derived from the PDFs. Through experiments using real sound data, we have confirmed that our framework is effective to generate permutation-free unmixed results.

Blind vector deconvolution: convolutive mixture models in short-time fourier transform domain

For short-time Fourier Transform (STFT) domain ICA, dealing with reverberant sounds is a significant issue. It often invites a dilemma on STFT frame length: frames shorter than reverberation time (short frames) generate incomplete instantaneous mixtures, while too long frames may disturb the separation. To improve the separation of such reverberant sounds, the authors propose a new framework which accounts for STFT with short frames. In this framework, time domain convolutive mixtures are transformed to STFT domain convolutive mixtures. For separating the mixtures, an approach of applying another STFT is presented so as to treat them as instantaneous mixtures. The authors experimentally confirmed that this framework outperforms the conventional STFT domain ICA.