Automatic Recognition of Speech in Stressful Environments
HUMAN ENGINEERING LAB ABERDEEN PROVING GROUND MD
Pagination or Media Count:
This document describes a body of research conducted for the purpose of improving the performance of automatic speech recognition ASR systems in noisy, stressful environments. To accomplish this, several avenues of basic research were investigated. First, novel techniques for noise reduction were developed. Second, three novel robust recognition systems were developed. One of these techniques relied on noise removal, another on linear least squares system identification, and the third on robust distance measures. Third, methods for characterizing stressed speech parametrically were developed. Among the descriptors used were various prosodic features, spectral features, and glottal excitation characteristics. Fourth, methods were developed that were shown to improve ASR performance in noisy or stressed conditions.