Problems in Blind Separation of Convolutive Speech Mixtures by Negentropy Maximization
Rajkishore Prasad, Hiroshi Saruwatari, Kiyohiro Shikano · 2003
ABSTRACT: This paper aims to examine suitability of the marginal statistics based contrast function e.g. negentropy for the separation of convolutive speech mixtures picked up by a linear microphone array. For this study we choose our frequency domain fixed-point ICA algorithm, based on negentropy maximization of the independent components. This algorithm is based on the heuristic assumption, in accordance with the Central Limit Theorem (CLT), that the gaussianity of mixed speech signal is more than that of unmixed individual. This assumption is true for long segment of speech in the time domain and the same is expected to hold even for small segments of speech in time domain and for every spectral bin for frequency sub-banded speech. In this paper we examine this assumption by estimating spectral kurtosis on the frequency time series of the signal obtained by taking Discrete Fourier Transform (DFT) of quasi-stationary segments of speech. It has been found that in more than 35 % of the frequency bins, speech signal fails to comply the CLT assumption, which in turn badly affects the separation performance of the fixed-point algorithm. 1.