Lexical stress estimation and phonological knowledge

Abstract

It is argued that the prosodic feature stress is useful in constraining the number of hypotheses a speech recognition system produces. A probabilistic algorithm is described for the estimation of the lexical stress pattern of English words from the acoustic signal using hidden Markov models (HMMs) with continuous asymmetric Gaussian probability density functions. Adopting binary (stressed or unstressed) syllable models, two five-state HMMs of the left-to-right type were generated, one for each value of the binary opposition. Training observation vectors were extracted from a corpus of bisyllabic stress-minimal word pairs, where each word occurred in a continuously spoken sentence. The vectors consisted of nine acoustic measurements based on fundamental frequency, syllabic energy and coarse linear prediction spectra. Evaluation of both stressed and unstressed models using a new set of recordings of the same word pairs yielded an average syllable-stress recognition rate of 94%.