Properties of Stochastic Perceptual Auditory-event-based Models for Automatic Speech Recognition

Recently, physiological and psychoacoustic studies have uncovered new evidence supporting the idea that human auditory processes focus on the transitions between spoken sounds rather than on the steady-state portions of spoken sounds for speech recognition. Stochastic Perceptual Auditory-event-based Models (SPAMs) were developed by Morgan, Bourlard, Hermansky and Greenberg to take this new evidence into account for word models in speech recognition by machines. This paper details our efforts to build a speech recognition system based on some of the properties of SPAMs. Although not all aspects of the complete SPAM theory have been implemented, we did find that fairly good recognition is possible with a system that concentrates almost exclusively on the transitions between speech sounds. Additionally, we found that such a system enhanced the more conventional phoneme-based system, which emphasized recognition of steady-state sounds. This blended system performed better than either system alone, especially in the case of noise-obscured speech.