Modeling the perception of features in the identification of English consonants

Abstract

The perception of voicing, place, and manner was investigated in a series of gating experiments in which subjects identified eight English consonants in six different duration conditions [T. L. Doeleman, J. Acoust. Soc. Am. 101, 3111(A) (1997)]. Subjects heard either all eight stimuli or two mutually exclusive subsets such that certain feature information was given (e.g., manner given in sets [p, t, b, d] and [f, s, v, z]). Although all subjects listened to the same 48 stimuli, identification significantly improved for subjects given place information (i.e., sets of labials and alveolars). The results are modeled using a spreading-activation connectionist model, the Integration-Competition model [Spivey-Knowlton (1996)]. In this model, the phoneme and feature representations compete with one another such that featural distinctions correspond to different weights assigned to the connection between feature and phoneme layers. The target phoneme has high activation from all featural arrays, resulting in the highest activation in the phoneme array. The increase in identification scores across duration conditions is modeled by the increase in activation of the target phoneme with successive time steps. The model also predicts observed consonant confusion patterns in that distracter phonemes sharing more features with the target phoneme will have higher activation levels.