A TF distribution for disturbed and undisturbed speech signals and its application to noise reduction. (Q1575755)

All quadratic time-frequency (TF) representations that provide the TF-shift covariance property belong to the well-known Cohen class. In recent years, several publications have dealt with the problem of choosing the kernel function with respect to the signal to be analyzed. This paper investigates a solution for disturbed and undisturbed speech signals aimed at achieving a good TF analysis for the noise reduction problem. First, we introduce a speech model that defines a certain class of signals. Based on the speech model we develop a new compound kernel consisting of the Margenau-Hill and the smoothed pseudo-Wigner distribution and demonstrate its superior performance compared with some popular distributions (Zhang-Sato distribution, pseudo-Wigner distribution, smoothed pseudo-Wigner distribution, S-method). Finally, the so-called `smoothed Margenau-Hill distribution' is used to design a time-frequency filter for the noise reduction algorithm based on the time-variant Wiener filter, for which we obtain promising results.