@InProceedings{Rykova2026_1287,
author = {Eugenia Rykova and Tanja Rinker and Angela Grimm},
booktitle = {Studientexte zur Sprachkommunikation: Elektronische Sprachsignalverarbeitung 2026, Tagungsband der 37. Konferenz},
title = {ASR-based Automatic Assessment of Oral Production Tasks in Multilingual Children},
year = {2026},
editor = {Günther Wirsching},
month = mar,
pages = {128--134},
publisher = {TUDpress, Dresden},
abstract = {The research project SPEAK aims to standardise a German language development test battery for multilingual children. This study evaluates the feasibility of automating a phonological production (nonword repetition, NWR) task and a vocabulary production task (Cross-linguistic Lexical Task, CLT) with the help of automatic speech recognition (ASR). The recommended accuracy threshold for ASR application in speech and language therapy is 80%. Five ASR models were tested with 858 audio recordings from the NWR task and 1267 audio recordings from the CLT noun production in German. Both tasks were administered to multilingual children and coded manually. The error rates (ERs) were calculated between the ASR transcriptions and the target, and the respective automatic phonemic transcriptions. For the CLT, the responses with the ER below 0.35 were accepted. For the NWR task, the transcriptions were post-processed to mimic the manual assessment process, and only the answers with ER = 0 were accepted. The mean accuracy of the automatic assessment was 69.7% for the NWR dataset, and 92.7% for the CLT dataset. High accuracy scores suggest the suitability of the ASR-based automatic assessment for evaluating multilingual children’s performance in the noun-production task. The accuracy scores in the NWR task are still too low. Importantly for language diagnostics, the greatest drawback of automatic assessment (in both tasks) is misrecognising correctly produced items, which would not lead to underdiagnosis of children with language difficulties. This result would be in line with the proposals to accept overdiagnosis rather than underdiagnosis.},
isbn = {978-3-95908-834-3},
issn = {0940-6832},
keywords = {Speech Analysis II},
url = {https://www.essv.de/pdf/2026_128_134.pdf},
}