Neural network based digit recognition system for voice dialling in noisy environments

During this decade voice dialling has became a potential feature to be implemented, e.g., in the mobile phones, but the techniques still suffer from defects of current algorithms when a real background noise is present. In this paper, a hybrid speech recognition system is discussed considering both...

Full description

Saved in:
Bibliographic Details
Published in:Information sciences 1999-12, Vol.121 (3), p.171-199
Main Authors: Salmela, Petri, Lehtokangas, Mikko, Saarinen, Jukka
Format: Article
Language:English
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:During this decade voice dialling has became a potential feature to be implemented, e.g., in the mobile phones, but the techniques still suffer from defects of current algorithms when a real background noise is present. In this paper, a hybrid speech recognition system is discussed considering both speaker-independent isolated and connected digit recognition in adverse conditions. Specifically, the recognition system consists of one or more multilayer perceptron networks (MLP) and hidden Markov models (HMM), which were trained using minimum classification error (MCE) method or embedded Viterbi (EV) training. The goal of the paper is to compare the performances of these two algorithms and study whether some advantage is gained in the MCE training by using a novel cost function, which puts more weight on the misclassified training samples than the conventional cost function (sigmoidal). The performance comparisons were made with a data set recorded in car in three different noise environments. The hybrid recognition system yield 98.18% and 74.45% accuracies for the isolated and the connected digits, respectively. The gradient equations are also derived for the MCE training considering multiple MLPs and the novel cost function in the recognizer structure.
ISSN:0020-0255
1872-6291
DOI:10.1016/S0020-0255(99)00077-8