Post-Nonlinear Sparse Component Analysis Using Single-Source Zones and Functional Data Clustering

TL;DR

This paper extends sparse component analysis to post-nonlinear mixtures using single-source zones and functional data clustering, enabling improved separation of speech signals with novel confidence measures and approximation techniques.

Contribution

It introduces a new PNL-SCA framework that employs weak sparsity, novel confidence measures, and functional data clustering, including extensions for general nonlinear mixtures.

Findings

Mutual information and LTSA-correlation outperform LTSA-PCA in detecting single-source zones.

Local linear approximation clustering is more accurate than B-spline methods.

The approach successfully separates speech signals in simulated PNL mixtures.

Abstract

In this paper, we introduce a general extension of linear sparse component analysis (SCA) approaches to postnonlinear (PNL) mixtures. In particular, and contrary to the state-of-art methods, our approaches use a weak sparsity source assumption: we look for tiny temporal zones where only one source is active. We investigate two nonlinear single-source confidence measures, using the mutual information and a local linear tangent space approximation (LTSA). For this latter measure, we derive two extensions of linear single-source measures, respectively based on correlation (LTSA-correlation) and eigenvalues (LTSA-PCA). A second novelty of our approach consists of applying functional data clustering techniques to the scattered observations in the above single-source zones, thus allowing us to accurately estimate them.We first study a classical approach using a B-spline approximation, and then two approaches which locally approximate the nonlinear functions as lines. Finally, we extend our PNL methods to more general nonlinear mixtures. Combining single-source zones and functional data clustering allows us to tackle speech signals, which has never been performed by other PNL-SCA methods. We investigate the performance of our approaches with simulated PNL mixtures of real speech signals. Both the mutual information and the LTSA-correlation measures are better-suited to detecting single-source zones than the LTSA-PCA measure. We also find local-linear-approximation-based clustering approaches to be more flexible and more accurate than the B-spline one.