Using Whisper LLM for Automatic Phonetic Diagnosis of L2 Speech: A Case Study with French Learners of English
Résumé
This paper reports on a pilot study to use Whisper’s large language model (LLM) as a tool for potential representation of segmental (phone) pronunciation errors. We compared the perfor- mance of the transcription outputs for the various models developed by the automatic speech recognition (ASR) system Whisper (Radford et al., 2022) ranging from 39 to 1,550 million parameters. We investigated 38 recordings of two paragraphs from Conrad’s Typhoon. The whisper transcriptions were compared to the original text that was read by these second-year French undergraduates. We used WER (Word Error Rate) and Levenshtein distance to assess the various graphic representations of Conrad’s reference text. We show how the differences can be transformed into operationalised feed- back for learners. We used expert phonetic knowledge to check the plausibility of the pho- netic interpretation with the signal (in particular the recall of H dropping produced by French learners). Our findings suggest that the tran- scriptions produced by the medium model con- verge with what a native speaker understands and that the tiny model produces alternate transcriptions that are plausible candidates for learner errors.
Origine | Fichiers produits par l'(les) auteur(s) |
---|