International Research Training Group
Language Technology
&
Cognitive Systems
Saarland University University of Edinburgh
 

On Nonlinear Transformations of Random Variables in Speech Feature Enhancement

Speaker: Friedrich Faubel

Abstract:

In model-based speech feature enhancement approaches, the distribution of clean speech is modelled as a Gaussian mixture, the distribution of background noise is modeled as a single Gaussian whose mean and covariance matrix are to be estimated from noisy speech recordings. The latter can be achieved with the expectation maximization algorithm; however, it requires approximating the distribution of noisy speech, based on the clean speech distribution, the previous "hypothesis" of the noise parameters and the relationship between clean speech, noisy speech and noise in the feature domain. As the relationship is non-linear, the problem devolves to approximating non-linear transformations of Gaussian random variables.

In this talk, we will present different methods for approximating the distribution of nonlinearly transformed Gaussian random variables, including a new one that we have recently developed.

Last modified: Fri, May 29, 2009 10:57:04 by

Valid HTML 4.01 Transitional Valid CSS!