We present the first automatic speech recognition (ASR) system specifically designed for Luganda-speaking individuals with dysarthria. We fine-tuned the wav2vec2-luganda model on a real dysarthric Luganda corpus of 6,347 utterances, augmented to 102,629 samples using 12 augmentation types simulating the physiological spectrum of dysarthric speech. Combined with a 5-gram KenLM language model trained on 50,515 Luganda sentences, our final system achieves 24.03% WER on the test set, a 73% relative improvement over the untrained baseline and the first sub-30% WER result for any Luganda dysarthric ASR system. This work was developed as part of Uganda's Inclusive Speech Technology Hackathon, hosted at Makerere University and delivered by Senses Hub in partnership with the Makerere Innovation and Incubation Center and UCL's Global Disability Innovation Hub, as part of the AT2030 programme co-funded by Google.org.
Mark Herman Nkugwa (Wed,) studied this question.