NVIDIA refines medical speech recognition

NVIDIA’s 28 October Nemotron 3.5 Transcribe release includes a recipe for Indian medical English. Training combines synthetic speech with Common Voice and MultiMed recordings, teaching written conventions for drug names, numbers and units.[1][2]

Generated speech uses pronounceable forms while labels retain the target spelling. On 2,764 clips from speakers excluded from tuning, NVIDIA reports word error rate falling from 22.44% for its best-prompted base model to 12.17%. Prompts and checkpoints were selected on a separate 855-clip subset.[2]

Brand-name drug matches still trailed the medical specialist Parrotlet; LibriSpeech test-clean word error rate rose slightly.[2]

What the test measures

EkaCare assembled its benchmark from internal staff, medical professionals and speakers at four medical colleges. The English collection contains 3,619 recordings spanning isolated terms, narrated sentences and conversations. These settings define the speech being tested.[3]

EkaCare’s evaluation work separately examines equivalent number and unit expressions and annotated medical terms. These semantic and keyword metrics help distinguish formatting differences from errors in the terms that carry a prescription’s meaning.[4]

MultiMed’s 2025 paper describes a different source of medical speech: professional YouTube channels. Lectures, interviews and other public formats supplied recordings, with transcripts reviewed by medical experts. Those materials bring varied voices and recording conditions into the corpus, while representing a different setting from private consultations.[5]

Report an error or suggest a correction