Voice AI Models Struggle to Pronounce New Pharmaceuticals

The DOSE (Drug-name Oral Synthesis Evaluation) report highlights a critical gap in clinical-grade voice technology. Researchers tested nine commercial systems against 274 medicines, including 146 names approved recently. Success rates for general-purpose models plummeted when moving from established drugs to newer ones. For instance, ElevenLabs' eleven_v3 model saw its accuracy drop from 93.0% on legacy names to 67.1% on new entries, while Google’s Gemini TTS fell from 89.1% to 61.6%.

Rajashekar Vasantha, CTO of Synthio Labs, noted that sounding human does not equate to clinical accuracy. Mispronunciation risks exacerbating existing patient safety issues related to sound-alike drug names, a category already monitored by the FDA and WHO. The study suggests these errors occur because training data often lacks coverage for the most recent pharmaceutical molecules. Synthio Labs’ own specialized model, RxPronounce, outperformed the group, maintaining an 87.0% success rate on newly approved names. The full dataset is now available via Hugging Face for developers seeking to improve safety benchmarks.