Integration across frequency bands for consonant identification

Abstract

A comparison of the predictions of models of integration to data on the reception of consonants filtered into a variety of frequency bands is reported. New data on the consonant identification are presented. Three experiments were conducted testing the following bands: experiment I, 0-2100 Hz and 2100-4500 Hz; experiment II, 0-700 Hz combined with 700-1400, 1400-2100, 2100-2800, and 2800-4500 Hz; experiment III, all combinations of 700-1400, 1400-2100, 2100-2800, and 2800-4500 Hz. The predictions of four models, Fletcher's [Speech and Hearing in Communication (Van Nostrand, New York, 1950)] independent errors model, Massaro's fuzzy logical model of perception [Proc. Int. Congress of Phonetic Sciences, Stockholm, Vol. 3, pp. 106-113 (1987)], and Braida's pre-labelling and post-labelling models of integration [Q. J. Exp. Psychol. A 43, 647-677 (1991)], were compared in terms of their ability to predict combined-band scores. At least two models were capable of predicting performance for each combined-band condition. For experiment I, all models were able to make satisfactory predictions. For experiment II, a variant of the pre-labelling model was able to make satisfactory predictions. For experiment III, no model was able to make satisfactory predictions, but the fuzzy logical model of perception and a variant of the pre-labelling model made relatively good predictions. Thus the ability of the models to predict performance depended more on whether the condition included the lowest frequency band than on the adjacency or frequency separation.

Perceptual Confusions Among Consonants, Revisited—Cross-Spectral Integration of Phonetic-Feature Information and Consonant Recognition

“…In order to prevent listeners from using information outside the slit, many studies (e.g., [14] and [16]) add noise surrounding the pass-band. Background noise is often used in psychoacoustic studies to preclude "off-frequency" listening in signal-detection tasks.…”

“…In this section, we analyze how well five stimulus-integration models (Flet1, Flet2, Pre-labeling, Post-labeling, and Fuzzylogic, as described in [16]), account for the cross-spectral integration data presented above. By doing so, we show that the conventional methods for modeling cross-spectral integration fail to predict the pattern of phonetic-feature decoding errors observed in the current study.…”

“…The Flet2 model is a variant of Flet1, in which error probabilities are determined separately for each consonant rather than being averaged across all consonants. The Flet1 and Flet2 models are described in more detail in [16].…”

“…Ronan and colleagues [16] evaluated several of these with respect to their ability to predict consonant-recognition accuracy, (i.e., percent correct). These models-known as "Pre-labeling" [17], "Post-labeling" [17], and Fuzzy-logic [18]-were originally designed to account for cross-modal integration in audio-visual speech recognition (e.g., [19]).…”

“…The Fuzzy-logic model assumes optimal integration is based on a Euclidean centroid in a Fuzzy-logic response space. In [16], subjects were asked to identify consonants passed through either single, band-pass filters (whose center frequencies ranged between 700 and 2400 Hz) or through a combination of these pass-bands. The three models' predictions for the combined band-pass conditions were based on subjects' responses to the single, band-pass-filtered consonants.…”