The problem of blind source separation of complex-valued sources from a linear mixture is addressed. We propose a deflationary algorithm for the sequential recovery of a set of communication signals, where each source is extracted by performing a Bounded Component Analysis of the linear mixture. The contribution of each recovered source to the observations is removed by minimizing its convex perimeter, without using second-order statistics. This implies to run a gradient descent algorithm several times. In order to accelerate the convergence, we have derived a fast step size that exploits the second-order information of the cost function by means of the augmented Hessian matrix. Computer simulations show that the proposed method is able to blindly separate even dependent sources, as long as they satisfy the BCA separability conditions. Also, the speed of convergence of this novel step size is compared with other classical approaches.
18 June 2015 in IEEE/ACM Transactions on Audio, Speech, and Language Processing
In this paper, we propose a method for solving the permutation problem that is inherent in the separation of convolved mixtures of speech signals in the time-frequency domain. The proposed method obtains the solution through maximization of a contrast function that exploits the similarity of the temporal envelope of the speech spectrum. For this purpose, the contrast calculation uses a global measure of similarity based on the recently developed family of generalized Alpha-Beta divergences, which depend on two tuning parameters, alpha and beta. This parameterization is exploited to best measure the similarity of the speech spectrum and to obtain solutions that are robust against noise and outliers. The ability of this contrast function to solve the permutation problem is supported by a theoretical study that shows that for a simple time-frequency speech model, the contrast value reaches its maximum when the estimated components are properly aligned. Several performance studies demonstrate that the proposed method maintains a high level of permutation correction accuracy in a wide variety of acoustic environments. Moreover, it produces better results than other state-of-the-art methods for solving permutations in highly reverberant environments.
Centroid-based clustering is a widely used technique within unsupervised learning algorithms in many research fields. The success of any centroid-based clustering relies on the choice of the similarity measure under use. In recent years, most studies focused on including several divergence measures in the traditional hard k-means algorithm. In this article, we consider the problem of centroid-based clustering using the family of -divergences, which is governed by two parameters, and . We propose a new iterative algorithm, -k-means, giving closed-form solutions for the computation of the sided centroids. The algorithm can be fine-tuned by means of this pair of values, yielding a wide range of the most frequently used divergences. Moreover, it is guaranteed to converge to local minima for a wide range of values of the pair ( ). Our theoretical contribution has been validated by several experiments performed with synthetic and real data and exploring the ( ) plane. The numerical results obtained confirm the quality of the algorithm and its suitability to be used in several practical applications.
19 June 2014 in IEEE Transactions on Neural Networks and Learning Systems
This paper studies the problem of the blind extraction of a subset of bounded component signals from the observations of a linear mixture. In the first part of this paper, we analyze the geometric assumptions of the observations that characterize the problem, and their implications on the mixing matrix and latent sources. In the second part, we solve the problem by adopting the principle of minimizing the risk, which refers to the encoding complexity of the observations in the worst admissible situation. This principle provides an underlying justification of several bounded component analysis (BCA) criteria, including the minimum normalized volume criterion of the estimated sources or the maximum negentropy-likelihood criterion with a uniform reference model for the estimated sources. This unifying framework can explain the differences between the criteria in accordance with their considered hypotheses for the model of the observations. This paper is first presented for the case of the extraction of a complex and multidimensional source, and later is particularized for the case of the extraction of subsets of 1-D complex sources. The results also hold true in the case of real signals, where the obtained criteria for the extraction of a set of 1-D sources usually coincide with the existing BCA criteria.
The blind decomposition of the observations, as a set of additive components of simpler structure, is a problem with many applications in scientific and practical fields. Our study assumes that the component signals are of bounded nature, and relies on the geometric decomposition of the convex set that supports the observations as a Minkowski direct sum of the convex sets that support the components. This last property, which is weaker than the mutual independence of the additive components of the observations, is sufficient for the essential identifiability of the bounded and indecomposable components. In practice, it is usual that the components lie in one-dimensional complex subspaces. Therefore, for this case, we describe a sequential method for their recovery.