Blind separation of speech mixtures based on nonstationarity

Abstract

This paper presents a method for blind separation of convolutive mixtures of speech signals, based on the joint diagonalization of the time varying spectral matrices of the observation records and a novel technique to handle the problem of permutation ambiguity in the frequency domain. Simulations show that our method works well even for rather realistic mixtures in which the mixing filter has a quite long impulse response and strong echos.

Speech intelligibility improvement using convolutive blind source separation assisted by denoising algorithms

“…The non-stationarity was first taken into account by Matsuoka et al (1995). This problem has been studied by Parra and Spence (2000a) and Pham et al (2003). It was shown that decorrelation is able to perform the BSS task for wide class of source signals.…”

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

“…In this paper, we only consider the permutation problem, where the order of the recovered source components at each frequency bin may not be consistent with each other (e.g. ½ [26,30] utilizes the continuity of the spectral components in adjacent frequency channels while [27,28] use direction of arrival estimation, [31] combines both previous techniques, and [32] utilizes statistical signal models. However, the performance of these algorithms can be degraded by acoustical noise.…”

“…For the FD-BSS, the joint approximate diagonalization method proposed in [30] is used for source separation at each frequency bin. Note that the same ICA algorithm was used here for the contrast methods [30,33] in our comparisons in Section 5.2.2.…”

“…For the FD-BSS, the joint approximate diagonalization method proposed in [30] is used for source separation at each frequency bin. Note that the same ICA algorithm was used here for the contrast methods [30,33] in our comparisons in Section 5.2.2. In addition, the scaling problem is addressed based on the minimal distortion principle [29] for all these methods before performing the permutation alignment.…”

“…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.…”