📄 Vowel Effects on L2 Perception of English Consonants by Advanced Learners of English

Summary

The paper asks whether the vowel following an English fricative or affricate changes how advanced Chinese learners map that consonant onto their native categories. It also asks whether the Perceptual Assimilation Model (PAM), which predicts difficulty from L1–L2 phonetic distance, explains these mappings. Sixty-seven university students (mostly Mandarin L1, some Cantonese L1, all daily Mandarin users with a decade or more of English study) heard roughly 70 words recorded by a native American English phonetician. The targets were /s, ʃ, ʒ, ʧ, ʤ, tr, dr/ placed before /i, æ, u/, and /t, d/ served as fillers. For each word, listeners picked the closest Mandarin onset from pinyin options and then rated how well it fit on a 0–5 scale. Overall, PAM-style distance predicted most mappings. However, several sounds shifted their preferred Mandarin match depending on the vowel, and some choices contradicted phonetic similarity. The author attributes these effects to coarticulation (gestural economy) and possibly to pinyin spelling. This matters because L2 perception models, and pronunciation teaching, need to account for context and articulation, not just static sound inventories.

Key points

• Participants: 67 biomedical students at a Guangdong university (56 Mandarin L1, 11 Cantonese L1). Their gaokao English scores ranged from 114 to 138 out of 150. They received 30 minutes of IPA training before testing.
• The task combined a forced choice among 11 pinyin onsets with a 0–5 Likert goodness rating. It was run through the wjx.cn phone survey system, with an 8-second response interval.
• Main mappings:
  • /ʃ/ was heard as "sh" 75% of the time, with a rating of 4.3.
  • /ʧ/ and /tr/ were mostly heard as "ch" (over 70%), rated 4.4 and 4.1.
  • /ʤ/ leaned toward "j" (55.4%), while /dr/ leaned toward "zh" (about 70%).
  • /ʒ/ split three ways: "r" 34.2% (rating 3.5), "y" 14.7%, and "zh" 11.8%.
• The /t, d/ control items were categorized correctly more than 96% of the time.
• Vowel effects:
  • /s/ was chosen as "sh" more often before /u/ (13 of 134 responses).
  • /ʃ/ drifted to "x" only before /i/ (14 of 170).
  • /ʒ/ favored "r" with /i/ and "y" with /u/.

Interesting ideas

• I1. Each trial records both a category choice and a separate fit score, so a confident match can be told apart from a reluctant one.
• I2. The overall consonant results looked tidy, and only splitting them by the neighbouring vowel revealed systematic drift.
• I3. Irrelevant control items with a known expected answer were mixed into the test to confirm the procedure itself works.
• I4. A small table sets the match predicted by the theory beside the observed rate to pinpoint exactly where the model fails, and only then proposes other causes such as spelling.
• I5. The author worried that the notation of the answer options (pinyin versus IPA) could itself bias choices through surface-form overlap with English spelling.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Witness parsers or the LLM teacher could attach a goodness rating to each silver label, so that silver data can be weighted by how well the chosen category fits rather than by vote count alone.
• I2 (tokens): Tagger and lemma error reports could be broken down by the immediate left/right context of each token, so that context-triggered confusions hidden in aggregate accuracy become visible.
• I3 (learn): Every evaluation batch could include trivially solvable control items, so that a drop on them flags a broken pipeline rather than a model weakness.
• I4 (tree): Parser evaluation could list cases where the predicted attachment disagrees with treebank gold together with a candidate explanation, which would make error analysis explainable and could seed repair rules.
• I5 (answer): School-test and QA readers could be checked for answers driven by surface string overlap between question options and text, rather than by the meaning graph.