Learning a Talker or Learning an Accent: Acoustic Similarity Constrains Generalization of Foreign Accent Adaptation to New Talkers

Abstract

Past research has revealed that native listeners use top-down information to adjust the mapping from speech sounds to phonetic categories. Such phonetic adjustments help listeners adapt to foreign-accented speech. However, the mechanism by which talker-specific adaptation generalizes to other talkers is poorly understood. Here we asked what conditions induce crosstalker generalization in talker accent adaptation. Native-English listeners were exposed to Mandarin-accented words, produced by a single talker or multiple talkers. Following exposure, adaptation to the accent was tested by recognition of novel words in a task that assesses online lexical access. Crucially, test words were novel words and were produced by a novel Mandarin-accented talker. Results indicated that regardless of exposure condition (single or multiple talker exposure), generalization was greatest when the talkers were acoustically similar to one another, suggesting that listeners were not developing an accent-wide schema for Mandarin talkers, but rather attuning to the specific acoustic-phonetic properties of the talkers. Implications for general mechanisms of talker generalization in speech adaptation are discussed.

Keywords: adaptation; foreign-accented speech; generalization; perceptual learning; spoken word recognition; talker specificity.

Figure 1

Experiment 1 (Multi 1 → Speaker 1) test results: Priming of /d/-final words (RT in fair-SEED trials minus RT in seed-SEED trials) and /t/-final words (RT in fair-SEAT trials minus RT in seed-SEAT trials) for participants exposed to critical words (Experimental group) or replacement words (Control group). Error bars represent standard errors of the mean.

Figure 2

Probability density plots of acoustic measures of exposure /d/-final words (vowel duration, closure duration, and burst duration) for the exposure talkers in Experiments 1 (black solid lines), 2 (black dashed lines) and 3 (black dotted lines), as well as the test talker (dark grey lines). For comparison, light grey lines show native-English token distributions. The area under each curve equals 1.

Figure 3

Experiment 2 (Speaker 2 → Speaker 1) test results: Priming of /d/-final words (RT in fair-SEED trials minus RT in seed-SEED trials) and /t/-final words (RT in fair-SEAT trials minus RT in seed-SEAT trials) for participants exposed to critical words (Experimental group) or replacement words (Control group). Error bars represent standard errors of the mean.

Figure 4

Experiment 3 (Speaker 4 → Speaker 1) test results: Priming of /d/-final words (RT in fair-SEED trials minus RT in seed-SEED trials) and /t/-final words (RT in fair-SEAT trials minus RT in seed-SEAT trials) for participants exposed to critical words (Experimental group) or replacement words (Control group). Error bars represent standard errors of the mean.

Figure 5

Test phase results across studies: talker-specific condition (Speaker 1 → Speaker 1, Xie et al., 2017), single-talker condition (Speaker 4 → Speaker 1, Experiment 3), multiple-talker condition (Multi 1 → Speaker 1, Experiment 1) and. Priming of /d/-final words (e.g., seed) and /t/-final words (e.g., seat) in the experimental group. In both related priming types, /d/-final words (e.g. seed) served as auditory primes. Error bars represent standard errors of the mean.