SCSMT: A Multilingual Children’s Speech Corpus for Singapore’s Mother Tongues

This paper presents the Singapore Children’s Mother Tongue Speech Dataset, a multilingual corpus containing around 30 hours of speech from 659 children aged 7-9 years across three languages: Mandarin, Malay, and Tamil. The dataset includes read speech from structured text and spontaneous speech from picture-based question responses. We detail corpus design, participants, data acquisition, preprocessing, and annotation procedures. Comprehensive phonological and semantic analyses were conducted to examine pronunciation patterns and response content quality. We identify common linguistic challenges specific to each language, providing insights for mother tongue education. This dataset is a valuable resource for children’s automatic speech recognition and speech assessment tools, while contributing to understanding multilingual language acquisition in young learners.