Text-Only Domain Adaptation for Low-Resource Medical ASR: Shallow Fusion and Density Ratio Methods for Hausa

Automatic speech recognition (ASR) systems often underperform on specialised medical vocabulary, particularly for low-resource languages where domain-specific audio data is unavailable. We study text-only domain adaptation for Hausa medical ASR, requiring no additional audio. We combine MMS-1B-FL102 with two language model (LM) fusion strategies, Shallow Fusion (SF) and the Density Ratio Approach (DRA), using a 4-gram KenLM trained on 19,421 Hausa health sentences and a 75,372-sentence general-domain Hausa LM. We evaluate 250 recordings of 25 physician-reviewed medical prompts by 10 speakers across six northern Nigerian states using prompt-level 5-fold cross-validation. SF reduces WER from 29.56% to 23.55% (20.3% relative), while DRA reduces it to 23.33% (21.1% relative). Both gains are significant (p<0.0001), while the SF--DRA difference is not (p=0.42). A general-domain LM alone reaches 23.96% WER, indicating that most of the improvement comes from external LM fusion, with the medical LM providing a smaller additional gain. Text-only LM adaptation can substantially improve Hausa medical ASR without domain-specific speech data. Code is available at https://github.com/Tripp808/hausa-medical-asr.