The effects of lexical neighbors on stop consonant articulation

Abstract

Lexical neighbors (words sharing phonological structure with a target word) have been shown to influence the expression of phonetic contrasts for vowels and initial voiceless consonants. Focusing on minimal pair neighbors (e.g., bud—but), this research extends this work by examining the production of voiced as well as voiceless stops in both initial and final syllable/word position. The results show minimal pair neighbors can result both in enhancement and reduction of voicing contrasts (in initial vs final position), and differentially affect voiced vs voiceless consonants. These diverse effects of minimal pair neighbors serve to constrain interactive theories of language processing.

Perception in a Variable but Structured World: The Case of Speech Perception

“…Table 4.2 shows the posterior expectation and 95% highest posterior density intervals for each of the prior belief parameters given the adaptation data above. One possible reason for this is that a substantial minority of English speakers produce pre-voiced /b/ (Lisker & Abramson, 1964;Goldrick et al, 2013), which is characterized by a lower (negative) VOT and a higher variance (often higher even than /p/). That is, across talkers, the /b/ VOT distribution parameters (mean and variance) have a bimodal distribution.…”

“…We assumed a single, unimodal prior distribution, and the prior beliefs we inferred to be most likely are consistent with a compromise between the two types of /b/ distributions that talkers actually produce. Figure 4.6 shows the bimodal distribution of /b/ VOTs observed by Goldrick et al (2013), which has one short-lag cluster around 10ms VOT with low variance, and another prevoiced cluster centered around -100ms VOT, with high variance. The model-inferred /b/ distribution is a reasonable compromise, lying in between these two clusters in its mean and variance.…”

“…First, existing studies of voiced stop production suggest that some talkers prevoice /b/ more than others, making the distribution of /b/ means across talkers bimodal (Lisker & Abramson, 1964;Goldrick et al, 2013). But these studies do not include enough talkers to properly assess whether this is really representative of populationlevel variability.…”

“…Moreover, the exposure distribution in Kronrod et al (2012) was flat over VOTs, providing little information about the underlying distributions. Given that short-lag /b/s appear to be more common across talkers (accounting for the majority of tokens in both Lisker & Abramson, 1964;and Goldrick et al, 2013), the best guess for a single, particular talker's /b/ distribution would be the short-lag distribution. A second, related reason is that if a talker produces any short-lag VOTs at all, the prevoiced cluster will not have a substantial effect on the optimal classification boundary, since it will have lower likelihood than either the short-lag or /p/ distributions for VOTs near the boundary.…”

“…This underestimates the actual level of variability across talkers. Based on production data from both read (Baese-Berk & Goldrick, 2009;Goldrick et al, 2013;Lev-Ari & Peperkamp, 2013) and conversational speech (extracted from the Buckeye corpus, Wedel in prep), actual cross-talker variability is on the order of a standard deviation of 3-4ms VOT for (short-lag) /b/ and 10-15ms VOT for /p/. This is likely due to the particular mathematical form we assumed listeners' prior beliefs to have.…”