Abstract
A feature transform algorithm is proposed using the maximum likelihood spectral transform method for robust speech recognition. The feature transform compensates for the mismatch between clean training data and noisy testing data. A new technique is proposed to estimate a transform matrix for the feature transform in the linear spectral domain with a small number of parameters, which allows fast nonlinear adaptation to be realised.
Original language | English |
---|---|
Pages (from-to) | 1313-1315 |
Number of pages | 3 |
Journal | Electronics Letters |
Volume | 40 |
Issue number | 20 |
DOIs | |
Publication status | Published - 2004 Sept 30 |
ASJC Scopus subject areas
- Electrical and Electronic Engineering