CoChat

The DOI matches its Crossref record; 67 of the 67 references it cites were found in a registry or an index; none of the 58 sources its claims cite is checked for support yet.

Unattended distributional training can shift phoneme boundaries

Open access
Chládková, Kateřina, Boersma, Paul, Escudero, Paola · 2022 · Bilingualism: Language and Cognition

Abstract

Listeners are sensitive to speech sounds’ probability distributions. Distributional training (DT) studies with adults typically involve conscious activation of phoneme labels. We show that distributional exposure can shift existing phoneme boundaries (Spanish /e/–/i/) pre-attentively. Using a DT paradigm involving two bimodal distributions we assessed listener's neural discrimination across three sounds, showing pre-to-post-test improvement for the two adjacent sounds that fell into different clusters of the trained distribution than for those that fell into one cluster. Upon unattended exposure to an intricate stimulus set, listeners thus relocate native phoneme boundaries. We assessed whether the paradigm also works for category creation (Spanish establishing a duration contrast), where it has methodological advantages over the usual unimodal-versus-bimodal paradigm. DT yielded a greater effect for the /e/–/i/ boundary shift than for duration contrast creation. It seems that second-language phoneme contrasts similar to native ones might be easier to acquire than new contrasts.

Overview

CoChat Summary

TL;DR

In a fully unattended EEG distributional-training paradigm, 40 native Spanish listeners were exposed to bimodal distributions shifted along either the native F1 /i/-/e/ dimension or a nonnative duration dimension. Training shifted the /i/-/e/ boundary pre-attentively, as indexed by the absolute mismatch response, but produced no significant comparable effect for duration contrast creation.

Key findings

  • For the absolute MMR, the predicted Boundary × Deviant interaction was significant for F1: +1.718 μV, t(35.999) = +3.017, 95% CI 0.563–2.873 μV, p = 0.0047; low-boundary training yielded a larger MMR for D1 than D2, and high-boundary training yielded a larger MMR for D2 than D1.
  • The Dimension × Boundary × Deviant interaction was +2.519 μV, t(35.999) = +3.129, 95% CI +0.886–+4.153 μV, p = 0.0035, indicating the effect was larger for F1 than for duration.
  • For duration, the Boundary × Deviant interaction was −0.801 μV, t(35.999) = −1.412, 95% CI −1.956–+0.354 μV, p = 0.17, nonsignificant and opposite to the predicted direction.
  • No learning effects were detected in the negative MMN: the predicted Boundary × Deviant interaction was −0.263 μV, p = 0.79, and its interaction with Dimension was +0.329 μV, p = 0.86; effects emerged only for the absolute MMR.

Methods

  • Participants were 40 native Spanish speakers (20 women, 20 men), 19–36 years old, functional Spanish monolinguals with low English proficiency, in the Netherlands less than 2 weeks; they were randomly assigned to the F1 or duration dimension and to low- or high-boundary training (10 per cell). Three additional people were excluded (2 technical error, 1 noisy EEG with >65% artifact-contaminated epochs).
  • Training was a 9-minute unattended exposure to 600 acoustically different synthetic vowels sampled from bimodal distributions (two equal-variance Gaussians, equal-area method, randomly permuted). F1 low boundary was 341 Hz and high boundary 400 Hz; duration low boundary was 136 ms and high boundary 164 ms. Native Spanish /i/-/e/ boundary is roughly 370 Hz; the duration boundary is roughly 150 ms.
  • Pre- and post-test used an unattended EEG oddball mismatch-response paradigm while participants watched a muted movie with Spanish subtitles; EEG was recorded from 64 active electrodes plus 7 external channels at 8 kHz and downsampled to 512 Hz. F1 test stimuli were standard 370 Hz, D1 314 Hz, D2 423 Hz; duration test stimuli were standard 150 ms, D1 124 ms, D2 181 ms. Analysis used linear mixed-effects models in R/lme4 on 720 data points (40 participants × 9 channels × 2 deviant types), with fixed effects Dimension, Boundary, Deviant, Laterality, and Anteriority, and participant random slope for Deviant.

Limitations

  • Stated: the authors cannot yet exclude that the F1 boundary shift is language-independent and applies universally to the frequency dimension; testing nativeness would require two continua A and B and native listeners of languages X and Y with/without native contrasts along them.
  • Stated: it is unknown to what extent Spanish listeners treated the isolated vowel stimuli as linguistic and to what extent the experiment mimicked a language-learning scenario; the distributions were atypical for Spanish, so the effect may reflect adaptation to a different language, accent, or atypical speaker.
  • Stated/evident: the duration effect was nonsignificant (p = 0.17), so the conclusion that category creation is smaller or absent is qualified by the usual caveat against interpreting nonsignificant p-values; the paper also notes no negative-MMN effects and only absolute-MMR effects, and that convergence problems forced a reduced random-effects structure with only a random slope for Deviant.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1017/s1366728922000086

Cites

67 of 67 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 (a) Typical unimodal and bimodal training distributions as used in the distributional training literature. The left and middle graphs show ' continuously sampled ' bimodal and unimodal training distributions such as those reported in Wanrooij and Boersma (2013), where the actual stimuli are represented by the vertical lines. The right graph illustrates the locations of the test stimuli traditionally used to assess the effects of distributional training. (b) Illustration of two bimodal training distributions (left and middle), and three test stimuli (right). The right graph illustrates that under the lowered-boundary distribution (solid curve), test stimuli Y and Z fall within the same peak, whereas test stimulus X falls into the other peak. In contrast, under the raised-boundary distribution (dashed curve), test stimuli X and Y fall within the same peak, whereas test stimulus Z falls into the other peak. (c) The actual training distributions used in the present experiment for the lowboundary (left graph, solid line) and high-boundary (middle graph, dashed line) training groups; for the meaning of D1, S, and D2, see text. The values of peak locations are marked on the x -axis for each training dimension (F1 and duration): top row of marks = F1 values in Hz; bottom row = duration values in ms. The right graph shows the stimuli used in the oddball paradigm in pre- and post-test, and their values in Hz or ms. The x -axis is scaled in ERB for F1 and logarithmically for duration.
Figure 1 (a) Typical unimodal and bimodal training distributions as used in the distributional training literature. The left and middle graphs show ' continuously sampled ' bimodal and unimodal training distributions such as those reported in Wanrooij and Boersma (2013), where the actual stimuli are represented by the vertical lines. The right graph illustrates the locations of the test stimuli traditionally used to assess the effects of distributional training. (b) Illustration of two bimodal training distributions (left and middle), and three test stimuli (right). The right graph illustrates that under the lowered-boundary distribution (solid curve), test stimuli Y and Z fall within the same peak, whereas test stimulus X falls into the other peak. In contrast, under the raised-boundary distribution (dashed curve), test stimuli X and Y fall within the same peak, whereas test stimulus Z falls into the other peak. (c) The actual training distributions used in the present experiment for the lowboundary (left graph, solid line) and high-boundary (middle graph, dashed line) training groups; for the meaning of D1, S, and D2, see text. The values of peak locations are marked on the x -axis for each training dimension (F1 and duration): top row of marks = F1 values in Hz; bottom row = duration values in ms. The right graph shows the stimuli used in the oddball paradigm in pre- and post-test, and their values in Hz or ms. The x -axis is scaled in ERB for F1 and logarithmically for duration.CC-BY · From the PDF
Image 2
CC-BY · From the PDF
Figure 3 The absolute difference waves at FCz for each dimension, training boundary location, and stimulus type. The MMR analysis window is shaded. Note that for the duration dimension, two windows are marked, because the analysis window differed between D1 (earlier window) and D2 (later window), see section 2.6.
Figure 3 The absolute difference waves at FCz for each dimension, training boundary location, and stimulus type. The MMR analysis window is shaded. Note that for the duration dimension, two windows are marked, because the analysis window differed between D1 (earlier window) and D2 (later window), see section 2.6.CC-BY · From the PDF

Verification

Claims the paper makes

0 Valid · 3 Not valid · 39 Not checked · 67 Own results and interpretations

47 more claims

Create a free account to see all 109 claims and to verify the 36 references still open.

See all claims — free account
  • It may state two separate claims as one. 35
  • Probably not a substantial claim. 38

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 67 references

Research works printed but not linked: 6

  1. Maye, 2001, BUCLD 25 Proceedings, 480
  2. Meister, 2011, Proceedings of 17th ICPhS, 1362
  3. Terry, 2015, Proceedings of the 18th International Congress of Phonetic Sciences
  4. Bohn, 1995, Speech perception and linguistic experience: issues in cross language research, 279
  5. Schreiber, 2013, Proceedings of Meetings on Acoustics ICA2013, 19, 060298
  6. Nixon, 2018, Interspeech 2018 – Proceedings of the 19th Annual Conference of the International Speech Communication Association, 1447

Cited by

See all 18

Metadata

Authors
Chládková, K., Boersma, P., & Escudero, P.
Affiliations
  • Czech Academy of Sciences (CZ)
  • University of Amsterdam (NL)
  • Australian National University (AU)
Corresponding author
Chládková, Kateřina
Venue
Bilingualism: Language and Cognition
Published
2022-03-25
Volume
25
Issue
5
Pages
827-840
ISSN
  • 1366-7289 (ISSN-L)
  • 1469-1841
OpenAlex
W4229046639
Cited by
18
References
67
Citations by year
2026: 2 · 2025: 3 · 2024: 4 · 2023: 5 · 2022: 3
FWCI
2.33
Type
journal article
Language
en
Open access
Locations
  • Bilingualism Language and Cognition (publishedVersion, cc-by)
  • UvA-DARE (University of Amsterdam) (publishedVersion, cc-by)
  • Data Archiving and Networked Services (DANS) (submittedVersion, other-oa)
Funders
  • Nederlandse Organisatie voor Wetenschappelijk Onderzoek (277-70-008, 446-14-012)
  • Univerzita Karlova v Praze (Primus/17/HUM/19)
  • Grantová Agentura České Republiky (18-01799S)
  • Australian Research Council (FT160100514)
Topics
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Phonetics and Phonology Research
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Categorization, perception, and language
  • Social Sciences › Psychology › Developmental and Educational Psychology › Language Development and Disorders
Keywords
phoneme boundaries, preattentive processing, bimodal distribution

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-19, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-19 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Unattended distributional training can shift phoneme boundaries. https://app.cochat.ai/papers/10.1017/s1366728922000086
The paper
Chládková, K., Boersma, P., & Escudero, P. (2022). Unattended distributional training can shift phoneme boundaries. Bilingualism: Language and Cognition. https://doi.org/10.1017/s1366728922000086

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.