Automatic Speech Recognition for Bilingual Children in Identification of Language Disorder

For the diagnosis and supervision of speech therapy, automatic detection of language abnormalities in children’s speech is crucial. Bilingual children confront a special problem when it comes to multiple diagnoses of developmental language disorder (DLD) because of their wide linguistic experience and expertise. Although there are presently not enough resources to enable it, dual-language testing offers the most precise categorization of DLD among bilinguals. This research examined whether multilingual children with DLD might be identified using dual-language automated speech recognition (ASR). The Bilingual English–Spanish Assessment—Middle Extension (BESA-ME) Morph Syntax was completed by 79 bilingual Spanish–English students in the second grade. Of these, 23 had confirmed DLD diagnoses and 56 did not. Their responses to some of the inquiries were scored programmatically by an ASR system the researcher built with a commercial speech-to-text technology, as well as manually by human examiners. The results showed that both procedures using the best-language score had similar levels of diagnostic accuracy and little overall agreement in item-by-item scoring. These findings show that when ASR is used to evaluate the students’ responses, the BESA-ME Morph syntax for bilingual second graders in Spanish and English is concurrently valid. More generally, this research offers early evidence that ASR is a multilingual expressive language evaluation method that is technically feasible.