MedASR

General speech recognition models often stumble over drug names, anatomy and radiology terms. MedASR was trained specifically on physician dictations, which gives developers a starting point that already understands much of the language clinicians use every day.

• Focuses on medical dictation across radiology, internal medicine and family medicine
• Uses a compact Conformer design of about 105 million parameters
• Can be run locally, deployed on Vertex AI or downloaded from Hugging Face
• Comes with Google notebooks for getting started and fine-tuning
• Feeds transcripts into models such as MedGemma for summaries or question answering

MedASR is best for healthcare software teams, researchers and startups building dictation, scribing or documentation tools who are comfortable working with models directly. A practical tip: test it on a sample of your own de-identified recordings before committing, and plan a fine-tuning pass if your users have strong accents, noisy rooms or unusual medication names, since Google's model card notes that performance varies on terms outside the training data. Always keep a clinician review step for any patient-facing records.