AI–HUMAN SCORE DISCREPANCIES IN AUTOMATED JAPANESE SPEAKING ASSESSMENT: THE ROLE OF ASR ERRORS
Article Information
Abstract
Automated speaking assessment offers a scalable alternative to
exclusively human scoring, but its reliability depends on multiple processing stages.
In transcript-based systems, automatic speech recognition (ASR) converts a
learner’s utterance into text before an artificial intelligence (AI) component assigns
a score. Errors in transcription may therefore influence the evidence available for
assessment. Previous research stresses automated scoring validation [1, 2]. This pilot
study examines score agreement and explores whether recognition errors plausibly
contribute to large differences.
Keywords
no

