Development and comparison of three syllable stress classifiers

The paper describes the development of three alternative techniques for the classification of syllable stress in fluent speech. They are based on: (1) neural networks that use contextual syllabic information; (2) first and second order Markov chains that depend on a new dynamic vector quantization approach; and (3) a rule based approach. Both the neural network and the statistical approach achieved performance above 80%, with the neural networks slightly outperforming the Markov models. Experimental results also show that stress classification could enhance speech recognition