AI Speech-Recognition Assessment of Oral Reading Fluency in isiXhosa

Early Grade Reading Assessments (EGRA) are one-on-one oral reading tests used to flag which children are falling behind in literacy, and are the standard tool in low- and middle-income countries, but they require trained human assessors and are rarely available in African languages beyond the few with existing speech-recognition support.

A research team led by the University of Cape Town, with the University of Pretoria, Western Sydney University, Universidade Federal de Pernambuco and Macquarie University, set out to automate EGRA scoring for isiXhosa, one of South Africa's official languages.

Over an 8-month data-collection effort supported by a Gates Foundation Grand Challenges grant, the team gathered 14,971 EGRA recordings, more than 19 hours of child speech, from grades 1-4 in South African schools; three independent human markers first validated the scoring approach against an EGRA expert, and fine-tuned speech-recognition models were then benchmarked against that human-labelled data.

The human markers reached 85% agreement with the EGRA expert, and the fine-tuned speech-recognition models reached roughly 91% diagnostic efficiency against the human-labelled data.

The published evaluation notes that no single model configuration minimised both false-positive and false-negative error rates simultaneously, meaning any early deployment would trade one kind of misclassification for another, and the work remains a research-stage validation rather than a fielded national assessment system.