Revisiting and Revising L2AM as a Model of Accent Speech Sounds Detection

This study examined how language familiarity and accent strength interference listener evaluations of non-native speech following our previously established L2 Accent Model (L2AM). Twenty-Seven Japanese participants rated speech samples in English, Chinese, and Japanese from speakers with varying accent strengths. Across all languages, weak-accented speech consistently received higher ratings across four perceptual categories: education, fluency, personality, and professionalism. Generalized linear mixed model analyses confirmed accent strength as a primary predictor of negative evaluations, even among intelligible speech. Findings also reveal that accent-based bias extends beyond Chinese listeners and varies with listeners' familiarity with the speech language. These results confirms and extends the model and highlight the role of accent as a social marker. Technically, the findings advocate for socio-linguistically informed speech recognition models that integrate subjective listener perceptions. The study underscores the need for fairer human and machine-based evaluations of accented speech in multilingual contexts.