Evaluating the Accuracy of Automatic Speech Recognition for Children’s Language Assessment

Language sampling is widely recognized as a robust method of evaluating and monitoring children’s language development, yet its use in clinical and classroom settings remains limited due to the time and expertise required to transcribe, annotate, and analyze samples. This study examined the accuracy of a child-trained automatic speech recognition (ASR) system to expedite transcription of narrative samples. Oral narrative retell samples from 443 children in their first year of school were transcribed using 3 methods: gold-standard transcription by trained speech-language pathologists; teacher-reviewed transcripts; and ASR-generated transcripts. Six language metrics were analyzed via Systematic Analysis of Language Transcripts, and statistical comparisons were made using generalized linear mixed models. Following management of transcription noise, ASR-generated transcripts produced comparable language metrics when compared with the gold standard. Teacher review added additional accuracy beyond ASR-generated transcription. Automated, accurate transcription of children’s language is increasingly feasible using child-trained ASR models.