Exploring Artificial Intelligence-Generated Transcripts for Supporting Students Who Are Hard of Hearing: A Preliminary Analysis of Transcription Errors in Subject Lectures

Automatic speech recognition (ASR) may help students who are hard of hearing review spoken explanations, but transcription errors could affect access to course content. This preliminary study examined ASR errors in three Korean EBS lectures (language, mathematics, and chemistry) and an English Technology, Entertainment, Design (TED) talk. Each lecture was played on an iPad in a quiet room and transcribed using Clova Note (Naver Corporation, Seongnam, Korea) and Plaud (Plaud Inc., San Francisco, CA, USA). The outputs were compared with reference transcripts using word error rate (WER) and character error rate (CER). For Korean, WER was calculated using space-delimited eojeols; for English, it was calculated using words. In the Korean language lecture, WERs were 20.27% for Clova Note (Naver Corporation) and 13.85% for Plaud (Plaud Inc.); in mathematics, 42.78% and 54.33%; and in chemistry, 28.62% and 28.44%, respectively. In the TED talk, WERs were 3.88% and 3.10%. In mathematics, the ranking of the two tools differed between WER and CER. Error rates varied across materials and measures, but these data cannot establish comparative performance across languages or benefits for students. Future studies should examine errors in key course content and evaluate how students use verified transcripts to understand lectures.