Badenhorst, Jacob AC2025-03-192025-03-192024-111865-09291865-0937https://doi.org/10.1007/978-3-031-78255-8_5http://hdl.handle.net/10204/14189Many children in South Africa cannot read for comprehension. This finding came from a recent international reading literacy study in which South Africa was placed last out of all 50 countries who participated [12]. During the study, it was pointed out that 75% of the learners participating came from disadvantaged backgrounds, a contributing factor to the overwhelming majority of Grade 4 learners that had already been deprived of proper school instruction. The development of automatic oral assessment applications could potentially alleviate this tragedy. A core component required by such foreseen software solutions is automatic pronunciation error detection (APED). Such APED capable of evaluating the speech produced by school learners in turn requires the development of local language child speech recognition (CSR) to succeed. CSR system development in the South African context may be attempted by refining adult pre-trained models. To this end, the paper provides a detailed breakdown of child data sets as well as reading assessment categories used to build and evaluate APED. While it is possible to select APED thresholds that do not penalise sentences read correctly, the more subtle reading disfluencies remained challenging to identify.FulltextenChild speech recognitionAutomatic pronunciation error detectionComputer assisted language learning ·Oral assessmentArticleN/A