READ: Developing a South African ART with Fine-tuned Mispronunciation Detection, Voice Detection, Multi-speaker Detection and Text Generation

Reading literacy is poor for most South Africans due to the numerous negative aftereffects of Apartheid. Both Phala et al. and Naidoo et al. refer to the Progress in International Reading Literacy Study (PIRLS) 2011, which revealed that 61% of school children were unable to read. PIRLS 2016 also showed that 78% of grade 4’s could not read for meaning, which is the aim of reading literacy (Phala and Hugo [2022]). Despite attempts to change this and a greater focus on reading literacy in government policy, there is still a reading literacy crisis in South Africa. According to the 2024 reports released by the Department of Basic Education (DBE) in South Africa, 80% of Grade 3 learners are not meeting grade-level reading requirements. This problem extends to grade 6, where 70% of students are reading below their grade level (Lucwaba [2025]).

One-on-one teacher-guided reading practice is a possible solution, but is impractical due to limited teaching time and large class sizes, therefore, the Automatic Reading Tutor (ART) provides a promising alternative. It has been found that ARTs can teach students similarly or even better than trained teachers (Mostow et al. [2003]). While sophisticated ART’s already exist, the software is often not freely available, making them impractical for large-scale use in South Africa

This is why our project aimed to use open-source resources to research and develop an ART, fine-tuned to recognize South African English accents in classroom environments and to generate reading materials based on student performance for young school children. The system was subdivided into the development of an improved ART system, a mispronunciation detection system for South African English accents, and a story generation system.