Gavin BROOKS    Department   Kyoto University of Foreign Studies  Department of British and American Studies, Faculty of Foreign Studies    Position   Associate Professor
Date	2023/09/10
Presentation Theme	Automated transcription and measures of LD in spoken texts
Conference	H-LRF Conference 2023
Promoters	H-LRF
Conference Type	International
Presentation Type	Speech (General)
Contribution Type	Collaborative
Country	Japan
Venue	Hiroshima University (Online)
Holding period	2023/09/10～2023/09/10
Publisher and common publisher	Gavin Brooks,Jen Jordan
Details	For L1 speech transcription, models such as OpenAI’s Whisper are becoming increasingly common. However, ASR for L2 English learners’ speech is challenging due to factors such as pronunciation errors, disfluency, and non-canonical grammar (Wang et al., 2021). The presenter examined the error rate in 100 samples of L2 learner-produced texts recorded in the classroom and noted that, although ASR accuracy is approaching that of human experts, there are still areas where it struggles. The results show that the accuracy of automatic transcription is becoming comparable to expert transcription, but they also highlight the parts where ASR-based transcription has difficulty, as well as best practices for recording and cleaning transcripts when using models such as Whisper to transcribe L2 presentations and discussions.