Time-domain fast fixed-point algorithms for convolutive ICA

Abstract

This letter presents new blind separation methods for moving average (MA) convolutive mixtures of independent MA processes. They consist of time-domain extensions of the FastICA algorithms developed by Hyvarinen and Oja for instantaneous mixtures. They perform a convolutive sphering in order to use parameter-free fast fixed-point algorithms associated with kurtotic or negentropic non-Gaussianity criteria for estimating the source innovation processes. We prove the relevance of this approach by mapping the mixtures into linear instantaneous ones. Test results are presented for artificial colored signals and speech signals.

Use of bimodal coherence to resolve the permutation problem in convolutive BSS

“…Convolutive BSS can be directly performed in the time domain [23][24][25] by deconvolution, but the computational complexity is high especially when the mixing filters have long taps. Based on the short-time stationarity of the speech signals and the linear time-invariance of the mixing process, an alternative is to perform convolutive BSS in the frequency domain by applying the short-time Fourier transform (STFT) to the observations.…”

“…Under the framework of independent component analysis (ICA) [17], the BSS problems have been extensively studied and many classical algorithms have been proposed for the instantaneous mixing model such as the ''J-H'' algorithm [18], JADE [19], Infomax [20], SOBI [21] and FastICA [22] algorithms. For the more complex convolutive mixing model, one can apply either the time domain deconvolution algorithms [23][24][25] or the frequency domain separation algorithms [12][13][14][15][26][27][28][29][30][31], which often suffer from the permutation and scaling ambiguity problems. Considering the bimodal nature of human speech, we could potentially improve the separation of the source signals from their audio mixtures utilizing the audiovisual coherence obtained by the integration of visual speech.…”

Audio-visual convolutive blind source separation

“…Convolutive BSS can be directly performed in the time domain [8] by deconvolution, but the computational complexity is very high and sometimes it cannot guarantee the convergence to a global optimum, especially when the mixing filters have long taps. Based on the short-time stationarity of the speech signal and the linear time-invariance of the mixing process, an alternative is to perform BSS in the time-frequency domain by applying the short-time Fourier transform (STFT) to the observations.…”

“…In equation (8), the direct summation of log-likelihood is very sensitive to outliers. It happens that one outlier may change the total summation greatly and result in a wrong decision.…”

Reverberant speech separation with probabilistic time–frequency masking for B-format recordings

“…To separate sources under reverberant environments, two types of methods are often used, namely time-domain (Aichner et al, 2002;Thomas et al, 2006;Nishikawa et al, 2003) and frequency-domain (Sawada et al, 2004;Araki et al, 2001;Saruwatari et al, 2001;Sawada et al, 2005) approaches, respectively. The time-domain methods are often based on the extension of the instantaneous ICA to the convolutive case, and the computational complexity associated with the estimation of the filter coefficients can be high, especially when dealing with the mixtures in a heavily reverberant environment, i.e.…”