Prior over Evidence: Stereotype-Driven Diagnosis in LLM-Based L2 Pronunciation Feedback

Large language models deployed for second-language pronunciation feedback rely on pretraining stereotypes rather than supplied speech evidence, according to an analysis of 1,800 utterances across three audio-capable models [1]. The study, posted to the arXiv preprint repository on 13 June 2026, tested the assumption that LLM diagnoses are grounded in the speech evidence provided to them [1]. Researchers evaluated 1,800 L2-Arctic utterances spanning six first-language backgrounds, four pronunciation dimensions, and five evidence conditions, from a text-only baseline to numeric acoustic features and raw audio [1]. Each cell was scored on rating accuracy against gold labels, evidence coherence assessing internal consistency, and grounded correctness evaluated against gold evidence [1]. A central finding was the decoupling of rating accuracy and grounded reasoning. Internally coherent reasoning supported a wrong rating in 39.6% of judged cells, while only 15.8% of cells showed coherent reasoning backing a correct rating [1]. This gap indicates that models can produce fluent, internally consistent justifications that are nonetheless incorrect [1]. Phoneme-level feedback converged to a fixed inventory of L2-English difficulty phones that recurred across all six L1 backgrounds and all evidence conditions [1]. The pattern suggests the models default to a stored stereotype of which sounds are hard for learners, regardless of the actual speech sample or the learner's native language [1]. Acoustic evidence improved ratings only when the supplied feature directly probed the target dimension. Textualised F0 range raised pitch-variation grounding from 0.18–0.19 to 0.45–0.62 across all three models, but the same audio waveform without textualised F0 values did not reproduce the improvement [1]. Stress and phoneme correctness, which require aligning a target pronunciation to its acoustic realisation, remained ungrounded [1]. The paper concludes that current general-purpose LLMs are more reliable as verbalisers of externally computed pronunciation evidence than as standalone diagnostic engines [1]. arXiv, which hosts the preprint, is an open-access repository of electronic preprints that are moderated but not peer-reviewed; it passed two million articles by the end of 2021 and now receives about 24,000 submissions per month [6]. Large language models are machine learning systems with many parameters, trained with self-supervised learning on vast amounts of text [8].