On the robustness of Speech Perception

Speech perception is remarkably robust despite many potential communicative obstacles such as background noise and talker variation. One reason for this is that speech sounds are selected to be auditorily distinctive even under adverse listening conditions. For many years, my colleagues and I have been exploring a general account of the distinctiveness of vowels and consonants referred to as the auditory enhancement hypothesis, which states that the phonetic properties of sound categories covary as they do largely because language communities tend to select properties that have mutually reinforcing auditory effects. Such relations of mutual enhancement create intermediate perceptual properties that serve, individually and in combination, as the basis for distinctive features. A second possible reason for the robustness of speech perception is currently being investigated in our laboratory. By comparing speech sound categorization performance of human listeners and Bayesian ideal classifiers, we have obtained evidence that human performance may be optimized with respect to the distributional properties—especially the degree of category overlap—of naturally produced speech sounds. Thus, perceptual robustness may be explained in terms of both auditory properties of speech sounds and strategies of listeners to optimize sound categorization.