YembaEGRA — Building an AI Speech-Recognition Corpus to Diagnose Reading Deficiencies in Cameroon's Yemba Language

Cameroon has more than 200 national languages, and primary schools routinely mix children whose reading proficiency should be assessed in a mother tongue that few trained evaluators speak; standard Early Grade Reading Assessment (EGRA) protocols require a trained assessor to individually test each child, which is slow and depends on scarce multilingual staff.

Researchers Arielle Kitio Tsamo, Paulin Melatagia Yonta and Sherelle Kana Azeuko at Université de Yaoundé I, working with EU Horizon 2020 MSCA-RISE 'Esperanto' project funding, built and published the YembaEGRA speech corpus (Mendeley Data, June 2025) for Yemba, a Bantu language spoken in Cameroon's Bamileke region: 69 level-3 pupils (36 girls, 33 boys) in Bamendou, Haut-Nkam division, each read 60 curriculum-linked words twice, yielding 120 labelled recordings per child, to train and validate automatic speech-recognition (ASR) models under a 'test-intervene-test again' diagnostic framework.

The 2024 paper proposes that an ASR model listens to a child read aloud, classifies which specific reading sub-skill (decoding, fluency, comprehension) is weak, and flags where a teacher should intervene; an educational facilitator selected the words to match eight thematic areas of the national curriculum, and a Yemba-language specialist translated and reviewed them.

As of publication, this is a corpus-building and model-development effort rather than a classroom-deployed and outcome-evaluated tool: no published data yet shows whether ASR-flagged interventions actually improve children's reading scores. The project is notable chiefly for demonstrating a replicable, low-cost method other under-resourced language communities could use to build their own EGRA-AI training data.