Did You Mispronounce or Did I Mishear? – Detecting Kanji Mispronunciations in Children's Oral Reading

Abstract

Oral reading of textbooks is a common practice in Japanese elementary education, yet it poses a unique challenge: kanji characters often have multiple readings, and without a listener to provide feedback, students may unknowingly reinforce incorrect pronunciations. To address this, we propose a novel algorithm for detecting kanji mispronunciations in children’s oral reading. Our method combines a deep learning-based automatic speech recognition (ASR) system with two probabilistic models: one modeling plausible kanji mispronunciations and the other modeling typical ASR errors. By aligning phoneme sequences generated from both the original text and the ASR output, the algorithm distinguishes genuine mispronunciations from transcription errors due to ASR. Experimental evaluation on speech data from children aged 6–9 shows that the proposed method successfully detects 84.6% of kanji mispronunciations that are included in the mispronunciation candidate dictionary of the probabilistic mispronunciation model.