We present methods to separate blindly mixed signals recorded in a room. The learning algorithm is based on the information maximization in a single layer neural network. We focus on the implementation of the learning algorithm and on issues that arise when separating speakers in room recordings. We used an infomax approach in a feedforward neural network implemented in the frequency domain using the polynomial filter matrix algebra technique. Fast convergence speed was achieved by using a time-delayed decorrelation method as a preprocessing step. Under minimum-phasemixing conditions this preprocessing step was sufficient for the separation of signals. These methods successfully separated a recorded voice with music in the background(cocktail party problem). Finally, we discuss problems that arise in real world recordings and their potential solutions. 1.

We present new results on the nonlinear PCA (Principal Component Analysis) criterion in blind source separation (BSS). We derive the criterion in a form that allows easy comparisons with other BSS and Independent Component Analysis (ICA) contrast functions like cumulants, Bussgang criteria, and information theoretic contrasts. This clarifies how the nonlinearity should be chosen optimally. We also discuss the connections of the nonlinear PCA learning rule with the Bell-Sejnowski algorithm and the adaptive EASI algorithm. Furthermore, we show that a nonlinear PCA criterion can be minimized using least-squares approaches, leading to computationally efficient and fast converging algorithms. The paper shows that nonlinear PCA is a versatile starting point for deriving different kinds of algorithms for blind signal processing problems.

Practical gradient-based adaptive algorithms for multichannel blind deconvolution and convolutive blind source separation typically employ FIR filters for the separation system. Inadequate use of signal truncation within these algorithms can introduce steadystate biases into their converged solutions that lead to degraded separation and deconvolution performances. In this paper, we derive a natural gradient multichannel blind deconvolutionand source separation algorithm that mitigates these effects for estimating causal FIR solutions to these tasks. Numerical experiments verify the robust convergence performance of the new method both in multichannel blind deconvolution tasks for i.i.d. sources and in convolutive BSS tasks for acoustic sources, even for extremely-short separation filters. 1.

In this chapter, we provide an overview of existing algorithms for blind source separation of convolutive audio mixtures. We provide a taxonomy, wherein many of the existing algorithms can be organized, and we present published results from those algorithms that have been applied to real-world audio separation tasks.

A new learning theory derived from the study of the dynamics of an abstract system of masses, moving in a multidimensional space under an external force field, is presented. The set of equations describing system's dynamics may be directly interpreted as a learning algorithm for neural layers. Relevant properties of the proposed learning theory are discussed within the paper, along with results of computer simulations performed in order to assess its effectiveness in applied fields.

In this paper sparseness measures are reviewed, extended and compared. Special attention is paid on measuring sparseness of noisy data. We review and extend several definitions and measures for sparseness, including the # , # norms. A measure based on order statistics is also proposed. The concept of sparseness is extended to the case where a signal has a dominant value other than zero. The sparseness measures can be easily modified to correspond to this new definition. Eight different measures are compared in three examples. It turns out that different measures may give complete opposite results if the distribution does not have a unique mode at zero. As conclusion, we suggest that the kurtosis should be avoided as a sparseness measure and recommend tanh-functions for measuring noisy sparseness.

This paper addresses a method of multichannel signal separation (MSS) with its application to cocktail party speech recognition. First, we present a fundamental principle for multichannel signal separation which uses the spatial independence of located sources as well as the temporal dependence of speech signals. Second, for practical implementation of the signal separation lter, we consider a dynamic recurrent network and develop a simple new learning algorithm. The performance of the proposed method is evaluated in terms of word recognition error rate (WER) in a large speech recognition experiment. The results show that our proposed method dramatically improves the word recognition performance in the case of two simultaneous speech inputs, and that a timing eect is involved in the segregation process. Indexing Terms: Blind signal separation, cocktail party speech recognition, dynamic recurrent networks, multichannel signal separation. submitted to Special Issue, Blind Si...

Source separation is becoming increasingly important in acoustical applications for spatial filtering. In the absence of any known source signals (blind case), a blind update equation similar to the natural gradient method [1] is presented, a derivative of which can be used in the case of known references (non-blind case). If some, but not all, source signals are known, blind-only algorithms are suboptimal, since some available information is not exploited. To overcome this problem, non-blind separation techniques can be incorporated. For the instantaneous mixing case (no time delays, no convolution), two different ways of combining blind and non-blind source separation methods are shown, namely an echo cancellertype and an equalizer-like approach. Simulations allow a comparison of the convergence time of both structures versus the convergence time of the blind-only case and clearly demonstrate the benefit of using combined blind/non-blind separation techniques.

A computationally simple nonlinearity in the form of a threshold device is shown to serve as contrast function in blind signal separation. Convergence is shown to be robust, fast, and comparable with that of more complex polynomial nonlinearities. Together with the known signum nonlinearity for super-Gaussian distributions, which basically is a threshold device with the threshold set to zero, the general threshold nonlinearity (with an appropriate threshold) can separate any non-Gaussian signals. 1. INTRODUCTION Blind signal separation using higher-order statistics either explicitly or implicitly has attracted many researchers whose main goal is to separate a set of mixed signals as fast as possible with the smallest residual mixing. Throughout this paper we assume a linear mixing and separation process as depicted in Fig. 1. A W s x u separation process mixing process sensors separated sources sources Figure 1: Blind source separation model. The measured signals x = [x 1 , . . . ...

We discuss the advantages of using overdetermined mixtures to improve upon blind source separation algorithms that are designed to extract sound sources from acoustic mixtures. A study of the nature of room impulse responses helps us choose an adaptive filter architecture. We use ideal inverses of acquired room impulse responses to compare the effectiveness of different-sized separating filter configurations of various filter lengths. Using a multi-channel blind least-mean-square algorithm (MBLMS), we show that, by adding additional sensors, we can improve upon the separation of signals mixed with real world filters. 1. INTRODUCTION Humans have the ability to focus their attention on any one sound in an environment filled with many different sounds. Digital audio systems, as well, would benefit from having this ability (termed by E. Collin Cherry in 1953 as the "cocktail-party effect."[3]); some potential applications include: instrument separation in a multitrack recording studio, s...