Lexical stress determination and its application to large vocabulary speech recognition

This study is concerned with the determination of lexical stress for isolated words from the acoustic signal. It is motivated by the suggestion that stressed syllables can provide islands of reliability where the phonetic information is more robust. In addition, stress information has been found to help lexical access. Our study consisted of three parts. First, we established through a lexical study that stress information can indeed provide strong constraints to be useful for large-vocabulary, isolated-word speech recognition. Second, an acoustic study was conducted to determine the acoustic correlates of lexical stress. Finally, a system was developed to determine the lexical stress patterns from the speech signal. The system performance, evaluated on some 1,600 words spoken by 11 talkers, indicated that the stressed syllable can be detected with a 2% error rate. The entire stress pattern can be determined with an error of 13%, with a significant amount of the error attributable to the system's failure in identifying the proper number of syllables.