An adaptive blind beamformer with an integrated single-channel noise reduction method for robust realtime blind speech extraction

Abstract

The performance of single-channel temporal noise reduction methods generally deteriorate in high noise environments, whereas spatial beamformers can maintain some level of speech enhancement. This paper presents a solution where a low complexity single-channel noise reduction method is integrated into the feedback control loop of an adaptive blind beamformer with the purpose of robust blind speech extraction in high noise environments. The proposed combined system outperforms each of the individual methods with respect to signal-to-interference ratio improvement for a wide range of operating conditions, and where the loss in estimated perceptual speech quality due to the combined system is tolerably low. Furthermore, the excess processing load in a hardware solution is comparatively insignificant for the proposed extended approach.

Statistical Analysis of a Local Quadratic Criterion for Blind Speech Extraction

“…In [10]- [12], a local quadratic criterion was derived for the purpose of Kurtosis maximization in a speech extraction application. The criterion was derived based on the inspiring paper by Yang regarding an iterative procedure for subspace tracking [16].…”

“…Following [10]- [12], this Kurtosis expression can be approximated and rewritten into a quadratic form (2) This expression is clearly quadratic with regard to . In this expression, the notation refers to the complex conjugate, and takes the real part of its argument and it is employed in order to ensure that the approximative Kurtosis expression is real valued.…”

“…However, the FastICA-like methods require a number of simplifications inside the optimization routine which make their approximations inexact at an optimum solution. A recently proposed technique approximates the Kurtosis value using a local quadratic function which is then solved at each iteration [10]- [12]. Robust methods for solving a quadratic problem exist in a rich variety today (see, e.g., [2]) which makes the proposed method practically tractable.…”

“…Robust methods for solving a quadratic problem exist in a rich variety today (see, e.g., [2]) which makes the proposed method practically tractable. Furthermore, the quadratic approximation of Kurtosis has been analyzed empirically in the past [10]- [12] and this method has shown itself capable of delivering a high level of speech extraction in various adverse environments. However, to date, no statistical analysis has been carried out for this particular approximation technique.…”

Complex-Valued Independent Component Analysis for Online Blind Speech Extraction

“…This approach is explored through a real-time speech enhancement application. It should be noted that in BSE, it is desirable to extract a dominant speech source (or a group of dominant sources) from an observed mixture of many sources [3]- [7].…”