Browse Prior Art Database

Model and Training Algorithm for Nonlinear Noise Removal

IP.com Disclosure Number: IPCOM000110417D
Original Publication Date: 1992-Nov-01
Included in the Prior Art Database: 2005-Mar-25
Document File: 2 page(s) / 72K

Publishing Venue

IBM

Related People

Das, S: AUTHOR [+4]

Abstract

Noise removal is accomplished by a mapping y=T(x). Ambient noise N is combined with training speech X to simulate a noise-corrupted speech signalY. After standard signal processing, splicing and projection, the resulting triples of frames (X,N,Y) are used to estimate the conditional expectation T(y)=E(X Y=y). The mapping T: Y --> X is based on a joint mixture model (X,Y,I) where the mixing variable I indexes classes determined by both the noise and the signal.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Model and Training Algorithm for Nonlinear Noise Removal

       Noise removal is accomplished by a mapping y=T(x).
Ambient noise N is combined with training speech X to simulate a
noise-corrupted speech signalY.  After standard signal processing,
splicing and projection, the resulting triples of frames (X,N,Y) are
used to estimate the conditional expectation T(y)=E(X Y=y).  The
mapping T: Y --> X is based on a joint mixture model (X,Y,I) where
the mixing variable I indexes classes determined by both the noise
and the signal.

      We seek a noise removing transformation
x = T(y)
for making noisy speech look like the training speech recorded in a
quiet environment.  Previous methods for this that we are aware of,
such as spectral subtraction, have been linear in the spectrum.  We
develop a natural nonlinear map which exploits the availability of
the ambient noise signal before and between segments of speech.

      Let I with values i e {1, 2,...,k} denote the random class
index corresponding to different types of noise and different types
of speech signal.  Denote by X a training vector and by Y an
artificially corrupted version of it.  Let N denote the corrupting
noise vector.  We approximate the joint distribution of (X,Y) by a
Gaussian mixture: given I=i the joint density is a 2d dimensional
multivariate gaussian density g(x,y ni, Si) with mean vector

                            (Image Omitted)

and covariance matrix
Let pi...