A fixed point solution for convolved audio source separation

Abstract

We examine the problem of blind audio source separation using Independent Component Analysis (ICA). In order to separate audio sources recorded in a real recording environment, we need to model the mixing process as convolutional. Many methods have been introduced for separating convolved mixtures, the most successful of which require working in the frequency domain [l], [2], [3], [4]. This paper proposes a fixed-point algorithm for performing fast frequency domain ICA, as well as a method to increase the stability and enhance the performance of previous frequency domain ICA algorithms.

Application of Geometric Dependency Analysis to the Separation of Convolved Mixtures

“…We begin with an overview of a typical frequency-domain approach to the separation of convolved mixtures, (see, e.g., [2] for more details) emphasing how the entire system can be understood as a composition of constrained sparse matrices followed by a partition of the resulting components into two subspaces.…”

“…In this application, we aim to place ones and zeros on the diagonal of H in order to select only those components which belong to a particular source. Given a partitioning of the components into two groups, we can therefore define two complementary ICA domain filters H (1) and H (2) to reconstruct the sources.…”

“…Each source was reconstructed as a stereo signal (using a Hanning window to recombine overlapping frames) and compared with those obtained from a frequency domain algorithm [2] using the same frame size of 256 samples from each channel. At a sampling rate of 16 kHz, these frames are only 16 ms long-much shorter than the room's impulse response-which places limits on the potential performance of both systems.…”

Using information theoretic distance measures for solving the permutation problem of blind source separation of speech signals

“…The resulting approaches will be tested on different speech mixtures recorded in real environments with different reverberation times in combination with different ICA algorithms, such as JADE [25], INFOMAX [4,26], and FastICA [27,28].…”

Effect of Central Limit Theorem non-compliance on blind separation of speech by negentropy maximization

“…One of the examples of algorithms based on such cost functions and non-Gaussianization of the signals are fixed-point ICA by the kurtosis or negentropy maximization in [12] for the separation of the instantaneousmixture for real valued signal and in [13] for complex-valued signals; however, this al gorithm has no strategy for solving the problems of permutation and scaling arising in speech signal separation in the frequency domain. The fixed-point algorithm for audio source separation can be found in [14,15]. The fixed-point Frequency Domain ICA (FDICA) algorithm for audio source separation work on the Time-Frequency Series of Speech (TFSS), and thus assumes obeyance of CLT from the TFSS in each frequency bin.…”