Stochastic Perceptual Auditory-event-based Models (aka SPAM)

Many physiological and psychoacoustic studies have suggested that human auditory processes focus on change in speech rather than on regions with constant or slowly-varying spectral properties. This property may help to improve the robustness of human speech recognition in the presence of some kinds of acoustic interference. Stochastic Perceptual Auditory-event-based Models (SPAMs) were developed by Nelson Morgan, Herve Bourlard, Hynek Hermansky and Steve Greenberg to incorporate this perspective into word models for speech recognition by machines. In a nutshell, this approach ties together the statistics of non-onset portions of speech to focus modeling power on the onset decisions. Preliminary experiments by the Realization Group at the International Computer Science Institute in Berkeley, CA have shown that this approach, when used in combination with more conventional models, appears to provide improved robustness for automatic speech recognition in the presence of slowly varying additive noise.


What is spam? (courtesy Dan Garcia)

Su-Lin Wu - November 29, 1995