An Analysis of Word-Level Intelligibility in EFL Learners Using ASR

This study investigates word-level intelligibility in the speech of Turkish EFL learners using Automatic Speech Recognition (ASR) as a diagnostic tool. A total of 54 university students read 100 commonly mispronounced English words embedded in three reading sessions. Transcriptions generated by ASR were analyzed to determine intelligibility rates, error types, and substitution patterns. The overall intelligibility rate was 71.2%, with substantial variation across participants and target words. While several words involving phonemes absent in Turkish (e.g., /θ/, /æ/) were frequently unintelligible, others traditionally labeled as difficult proved to be intelligible. Error analysis revealed that the most frequent intelligibility failures were wrong-word substitutions, some of which carried serious semantic or social consequences. Despite the absence of feedback or instruction, intelligibility rates improved modestly across sessions, suggesting potential benefits from repeated oral production. These findings challenge the validity of static “difficult word” lists and support the use of ASR-based, context-sensitive approaches for pronunciation evaluation and pedagogy.