A method to identify noise-robust perceptual features: Application for consonant /t/

Abstract

This study focuses on correlating speech confusion patterns, defined as consonant-vowel confusion as a function of the speech-to-noise ratio, and a model acoustic feature (AF) representation called the AI gram, defined as the articulation index density in the spectrotemporal domain. By collecting many responses from many talkers and listeners, the AF and psychophysical feature (event) is shown to be correlated via the AI-gram model and the confusion matrices at the utterance level, thereby explaining the listener confusion. Consonant /t/ is used as an example to identify its primary robust-to-noise feature, and a precise correlation of the acoustic information with the listeners' confusions is used to label the event. The main spectrotemporal cue defining the /t/ event is an across-frequency temporal coincidence, wherein frequency spread and robustness vary across utterances, while the event remains invariant. The cross-frequency timing event is shown to be the key perceptual feature for consonants in a vowel following context. Coincidences are found to form the basic element of the auditory object. Neural circuits used for coincidence in binaural processing for localization across ears are proposed to be used within one ear across channels. It is further concluded that the event is based on the audibility of the /t/ burst rather than on any superthreshold property.