No Detail Found!

Towards discriminative training estimators for HMM speech recognition system. J. of

Author(s)
M. Frikha Ecole Nationale d’Ingenieurs de Sfax, ENIS’, Department of Genie Electrique, SFAX, Tunisia
Z. Ben Messaoud Ecole Nationale d’Ingenieurs de Sfax, ENIS’, Department of Genie Electrique, SFAX, Tunisia
A. Ben Hamida Ecole Nationale d’Ingenieurs de Sfax, ENIS’, Department of Genie Electrique, SFAX, Tunisia
Abstract
This study investigates the issue of improving the discriminative training capabilities in Hidden Markov Model (HMM) isolated word recognition task. Hence for, two optimization criterions in the training phase are focused, the minimization of recognition Word Error Rate (WER) according to the Baum-Welch based Maximum Likelihood Linear Estimation (MLE) and the Maximum Likelihood Linear Regression (MLIR) adaptation training criterion. For this purpose, the Statistical Learning Theory (SLT) and the MLIR adaptation are applied in order to analyze, in the sense of minimum word error rate, the consistency of the training estimator in clean and mismatched environmental conditions. Several experiments were carried out. They all aimed to find an efficient training estimator algorithm with good generalization property and allowing a good training error rate with a significant training data reduction. The obtained results show that it exists an optimal specified training conditions which should be reached in order to guarantee an optimal specified characteristics of the HMM based isolated word recognition system.
Publication Details
Page(s) 3891-3899
DOI DOI not available
Published Journal: Journal of Applied Sciences, Volume: 7, Issue: 24, Year: 2007

Request the full text from PASTIC

0

Citations

0

Downloads

36

Views