Robust detection of phonetic features incritical bands

We consider how to detect phonetic features in noisy bandlimited speech. We propose an automatic method based on the hypothesis that independent feature detectors, working in parallel, account for the robustness of auditory strategies. Our method consists of three stages: rst, speech is ltered into critical bands and enhanced by nonlinearities; second, pho-netic cues are derived from narrowband measurements of periodicity and signal-to-noise ratio; third, signals from diierent bands are combined to make a global decision. These stages are formulated as components of a probabilistic graphical model, represented by a multilayer Bayesian network. The binary hidden variables in the model indicate phonetic cues in diierent parts of the frequency spectrum. We apply the model to detecting the phonetic feature +=?sonorant] that distinguishes vowels, nasals, and approximants from stops, fricatives, and aaricates. Implications for automatic speech recognition are discussed.