Ghanaian language speech recognition improves adolescent healthcare
Summary and headline written by AI from the source article. How we work
Researchers evaluated automatic speech recognition (ASR) systems in Twi, Dagbani, and Ewe, three languages spoken in Ghana, to improve adolescent health communication.
They benchmarked five ASR systems using both general religious texts and a dataset focused on youth sexual and reproductive health. While larger models showed initial promise, the team found that fine-tuning a smaller, more efficient model. Qwen3-ASR-0.6B, with a substantial corpus of Ghanaian Bible recordings yielded the most significant gains.
This fine-tuning dramatically improved performance, especially for Ewe, reducing word error rate from 109.3% to 64.8%. The team then tested the technology with KasaHealth, a voice-first application delivering adolescent sexual and reproductive health information. Fifty community members using KasaHealth gave the application a 100% chat-approval rate and a 72% rating of “Good” or “Excellent” for translation quality.
The findings suggest that the biggest obstacle to accurate speech recognition in these languages is a lack of relevant training data, not the underlying technology itself. Further development relies on expanding these in-domain datasets to address gaps identified during real-world testing with KasaHealth.
