Deepgram has extended its Nova-3 Medical speech-to-text model to 10 languages (Dutch, English, French, German, Hindi, Italian, Japanese, Portuguese, Russian, Spanish) with support for batch, streaming, and automatic code switching via a single model and parameter set. Compared to its own general multilingual model, Nova-3 Medical Multilingual reduces word error rate by 25-30% and entity error rate (conditions, drugs, doses) by 37-40%. It also claims the lowest average batch WER among competitors including ElevenLabs Scribe v2 Medical and Microsoft MAI 1.5. Deepgram separately upgraded its English-only Nova-3 Medical model, cutting WER by 23% and EER by 22%, applied automatically to existing API calls with no code change required.
Table of contents
Medical Specialization That Holds Across LanguagesNot All Transcription Errors Are EqualCompetitive Accuracy Across the Full Language SetBuilt for Real-Time Medical ApplicationsAlso New: A Better Nova-3 Medical for EnglishHow It WorksStart Building with Nova-3 Medical MultilingualQuestions this post answers
Does Deepgram's Nova-3 Medical model support languages other than English now?
Yes, Nova-3 Medical Multilingual adds support for Dutch, English, French, German, Hindi, Italian, Japanese, Portuguese, Russian and Spanish in both batch and streaming, with automatic code switching between languages. It is used with a single parameter set: model=nova-3-medical and language=multi, and is also available for self-hosted deployments. Explore how daily.dev developers track speech-to-text model updates like this multilingual medical release.
How much more accurate is Nova-3 Medical Multilingual than a general multilingual speech model for medical transcription?
It cuts word error rate by 25% in batch (6.62% to 4.96%) and 30% in streaming (8.32% to 5.86%) versus Deepgram's general-purpose Nova-3 Multilingual model, averaged across 10 languages. Entity error rate, covering drugs, doses and conditions, drops even further: 37% in batch and 40% in streaming, with dose errors falling 64% in batch and 81% in streaming. Developers comparing speech recognition accuracy for healthcare apps can follow benchmarks like these on daily.dev.
Do I need to change my code to get the upgraded Nova-3 Medical English model on Deepgram?
No code change is required. Any existing call to model=nova-3-medical with an English language code automatically uses the upgraded model, which cuts word error rate by 23% and entity error rate by 22% in batch versus the previous English version, and reduces streaming WER from 4.35% to 3.36%. Teams upgrading API-dependent models without breaking integrations can track changes like this on daily.dev.
Share this post