ASR Domain Adaptation for Ghanaian Adolescent Health Communication

Researchers fine-tuned a Qwen3-ASR-0.6B model on a 90k-sample Ghanaian Bible corpus to improve speech recognition in Twi, Dagbani, and Ewe. The adaptation significantly reduced Word Error Rate, specifically dropping Ewe WER from 109.3% to 64.8%.