Text-based decoding cuts errors in a small Hausa medical speech study

Medical vocabulary poses a particular challenge for speech recognition when a language has little specialized audio available for training. In an arXiv preprint, Oche David Ankeli and colleagues investigate an alternative for Hausa: leave the speech model unchanged and use written text to guide transcription. Their results support external language-model assistance on a constructed benchmark, while showing that the specifically medical contribution is comparatively small.

Changing decoding, not the speech model

The researchers used MMS-1B-FL102 with its Hausa adapter as a fixed acoustic backbone. They did not fine-tune it on medical recordings. Instead, they trained a compact language model on 19,421 sentences collected from the health sections of BBC Hausa and Deutsche Welle Hausa.

This model uses four-word sequence statistics to score possible transcriptions. Its KenLM binary occupies 16.76 MB, and the authors report that training it took 2.5 seconds on a CPU. Those figures describe the language-model component, not the size or computational requirements of the complete speech-recognition system.

The study tested two ways of incorporating these text statistics. Shallow Fusion combines acoustic and medical-language-model scores during beam search, which considers alternative transcriptions. The Density Ratio Approach then rescores 50 candidates from the Shallow Fusion search, adding a medical-language score and subtracting a weighted general-language score. The intended correction is to reduce double-counting of language preferences already represented in the recognizer.

That general-language model was trained on 75,372 sentences from several public Hausa resources. It approximates the recognizer’s existing language prior rather than reproducing it exactly.

A controlled, limited speech benchmark

Ten volunteers from six northern Nigerian states each recorded 25 prompts on personal mobile phones in quiet indoor environments, producing 250 utterances. The prompts ranged from everyday health phrases to longer, medically dense sentences. Initial drafts were generated with a language model, then independently reviewed and corrected by two medical doctors for medical accuracy and natural Hausa phrasing.

Evaluation used five-fold cross-validation organized by prompt. In each fold, the researchers selected decoding settings using recordings of 20 prompts, froze those settings, and evaluated recordings of the remaining five. Every prompt therefore received a held-out prediction, but every speaker appeared in both tuning and evaluation portions. This tests performance on held-out prompts, not unfamiliar speakers.

The researchers estimated uncertainty by resampling whole prompt groups, keeping their ten speaker recordings together. That respects the repeated-prompt structure, although only 25 independent prompt groups were available.

Clear gains, little separation between methods

Word error rate, which counts word substitutions, deletions and insertions relative to the reference, was 29.56% for greedy decoding without an external language model. It fell to 23.55% with medical Shallow Fusion and 23.33% with density-ratio rescoring: relative reductions of 20.3% and 21.1%.

Both improvements over greedy decoding were statistically significant in the authors’ analysis, with p-values below 0.001. Density-ratio rescoring’s 0.22-percentage-point advantage over Shallow Fusion was not significant, however, with p = 0.423. It changed only 28 of the 250 Shallow Fusion transcriptions.

A crucial comparison limits the domain-adaptation claim. Shallow Fusion using only the general-language model achieved 23.96% word error rate, just 0.41 percentage points behind the medical model. Combining general and medical text yielded 23.85%, no improvement over medical-only Shallow Fusion. Most of the gain therefore came from adding an external language model at all.

Better term recall is not clinical validation

The authors also measured Medical Term Error Rate using a fixed lexicon of 66 term strings. Medical Shallow Fusion reduced this measure from 31.27% to 24.18%, but general-language fusion scored slightly better at 23.80%. The terminology results therefore do not isolate a distinct medical-model advantage.

This metric has important limits: it ignores term position, folds certain letter distinctions, treats morphological variants separately and does not penalize inserted medical terms. Part of its lexicon was also selected using the same language models being evaluated. Anatomy terms remained especially difficult, with errors worsening slightly under both medical fusion methods.

The broader limits are substantial. These were purpose-built sentences read aloud, not spontaneous consultations; speakers were mostly young adults, and several Hausa varieties were underrepresented. The authors release code and trained language models, but withhold evaluation audio for privacy. They explicitly describe the resources as research tools without clinical validation—not systems for unsupervised decisions or unreviewed medication and dosage transcription.