Published in Symposium on Machine Learning for Development 2024. This article was peer reviewed before publication.
Kigali Institute of Applied AI
Low-resource speech recognition for Kiswahili clinical dictation
Abstract
Clinical documentation in East African hospitals is time-consuming and often incomplete. We present a speech-recognition system for Kiswahili clinical dictation trained with only 40 hours of transcribed audio by combining self-supervised pretraining with targeted data augmentation. On a held-out set of ward-round recordings, the system reached a word error rate of 14.2 percent, a relative improvement of 38 percent over a supervised baseline. We analyse error patterns around code-switching between Kiswahili and English medical terms and describe a lightweight deployment that runs on commodity hardware without cloud connectivity.
Keywords: Speech recognition, Kiswahili, Clinical NLP, Low-resource
Cite as: Amina Yusuf, John Otieno, & Fatuma Ali (2024). Low-resource speech recognition for Kiswahili clinical dictation. Symposium on Machine Learning for Development 2024. https://quorumpress.org/papers/low-resource-speech-recognition-pz8as
© 2024 Yusuf, Otieno and Ali. Open access under CC BY 4.0.
Introduction
Automatic speech recognition can reduce the documentation burden on clinicians, but most systems are trained for high-resource languages. Kiswahili clinical speech, with frequent code-switching, is poorly served.
Approach
We fine-tuned a self-supervised acoustic model on 40 hours of transcribed ward-round audio, using augmentation to expand coverage of medical terminology and speaker variation.
Results
The system achieved a 14.2 percent word error rate, a 38 percent relative improvement over the supervised baseline. Most residual errors occurred at Kiswahili-English switch points involving drug names.
References
[1] A. Baevski et al., “wav2vec 2.0: self-supervised learning of speech representations,” 2020.
Cite this paper
Amina Yusuf, John Otieno, Fatuma Ali. Low-resource speech recognition for Kiswahili clinical dictation. Symposium on Machine Learning for Development 2024, 2024. https://quorumpress.org/papers/low-resource-speech-recognition-pz8as
Publication ID: pz8as