Automatic Assessment of Language Developmental Disorders in Non-English Contexts

Assessing the language development of children is crucial for early detection and intervention in language developmental disorders. However, it poses challenges due to the difficulty in gathering sufficient data to accurately evaluate children's language skills within limited medical consultation hours and in unfamiliar environments for children. While machine learning (ML) offers potential for efficient data collection in natural environments for children, ML models designed for specific clinical purposes are often trained on English and perform poorly in non-English contexts. In this study, we used widely available ML tools such as OpenAI's GPT and Azure's Speech-to-Text, which are trained with large datasets including diverse languages, to develop an automated pipeline for speaker identification, speech transcription, and speech content analysis tailored for assessing Japanese language development. We have demonstrated the effectiveness of our pipeline in everyday settings for children, including special educational centers and public nursery schools. Furthermore, we found that the vocabulary size of children as-sessed by our pipeline significantly correlates with developmental scales evaluated by teachers, indicating the potential value of our pipeline as a reliable assessment tool for children's language development in non-English contexts.