Processing reduced word-forms in speech perception using probabilistic knowledge about speech production.

Abstract

Two experiments examined how Dutch listeners deal with the effects of connected-speech processes, specifically those arising from word-final /t/ reduction (e.g., whether Dutch [tas] is tas, bag, or a reduced-/t/ version of tast, touch). Eye movements of Dutch participants were tracked as they looked at arrays containing 4 printed words, each associated with a geometrical shape. Minimal pairs (e.g., tas/tast) were either both above (boven) or both next to (naast) different shapes. Spoken instructions (e.g., "Klik op het woordje tas boven de ster," [Click on the word bag above the star]) thus became unambiguous only on their final words. Prior to disambiguation, listeners' fixations were drawn to /t/-final words more when boven than when naast followed the ambiguous sequences. This behavior reflects Dutch speechproduction data: /t/ is reduced more before /b/ than before /n/. We thus argue that probabilistic knowledge about the effect of following context in speech production is used prelexically in perception to help resolve lexical ambiguities caused by continuous-speech processes.Keywords: spoken-word recognition, continuous-speech processes, /t/-reduction, eye tracking, DutchContinuous speech is peppered with noncanonical pronunciations of words. Continuous-speech processes, such as assimilation, reduction, and deletion, operate during speech production and can substantially change the way words are realized. For example, a speaker of British English may say [waImbɔʔ1] in a casual request for a wine bottle, but say [waIn] and [bɔtəl] when asked how these words should be pronounced. It is the listener's task to recognize spoken words in spite of this variability. A particularly acute problem is when a continuous-speech process creates a lexical ambiguity. We examine that situation here, focusing on the effects of word-final /t/ reduction. If a speaker does not clearly pronounce the final [t] of duct, for example, how does the listener work out that they are not talking about a duck?Two answers to this question have been proposed. The first is based on lexical storage. There are in fact three radically different versions of this view. Lahiri and Marslen-Wilson (1991) and Lahiri and Reetz (2002) proposed a model of the mental lexicon in which entries are highly abstract and phonologically underspecified. A lexical entry for a given word only includes the phonological features that are reliably associated with tokens of that word. For instance, the alveolar nasal consonants in English are assumed to be specified only as nasal, with no specification of place of articulation, which may vary due to phonological assimilation. The token [waIm] can be recognized as an instance of the word wine because there is no mismatch between the last nasal segment [m] in the input and the lexical representation for this segment. Similarly, lexical entries may be underspecified for the presence of a word-final /t/. The input [d∧k] would thus match a stored lexical representation corresponding to duct. The second lexical-storage a...

Perceptual adaptation to segmental and syllabic reductions in continuous spoken Dutch

“…In the present experiments, we used a variant of the visual-world eye-tracking paradigm (Mitterer & McQueen, 2009b) that was modified to permit study of perceptual learning. Different groups of participants were exposed to one of two different forms of reduction, and then tested on both forms.…”

The mental lexicon is fully specified: Evidence from eye-tracking.

“…There are numerous problems with this argument. First of all, the standard error of the estimated (nonsignificant) effects were consistently small enough to lead to significant effects, given the effect sizes that have been observed for phonological effects in eye-tracking experiments (Gow & McMurray, 2007;Mitterer & McQueen, 2009 vealed phonological context effects that were not apparent in behavioral data (Mitterer & Ernestus, 2006;Mitterer & McQueen, 2009). Third, the current enterprise provided not one but four tests of the asymmetric-matching hypothesis.…”

“…To examine whether this standard error is excessively large and prevents us from finding an effect, it is useful to consider the effect sizes usually found in visual-world experiments testing phonological effects. Two papers examined context effects in the recognition of phonological variants with a similar eye-tracking paradigm (Gow & McMurray, 2007;Mitterer & McQueen, 2009). Because those papers investigated rival theories to the FUL model, it is reasonable to assume that effects of phonological underspecification should be of a similar size as the effects reported in these experiments.…”

“…This finding contrasts with assumption that the presence or absence of default features in the input does not influence the activation level of words with underspecified segments. Finally, context effects in the perception of /t/ reduction are also difficult to explain in the framework of the FUL model (Mitterer & Ernestus, 2006;Mitterer & McQueen, 2009).…”

“…To make the task less taxing, we made use of a recent innovation of the visual world paradigm. Mitterer and McQueen (2009) presented printed words with accompanying objects, which makes the instructions easier to follow. Figure 1 shows an example of such an experimental display.…”