Publication date: 18 September 2026
Deepgram released Nova-3 Pharma on 17 September 2026, adding a speech-to-text model tuned specifically for drug names and pharmaceutical vocabulary. It is available through the hosted API for English batch and streaming transcription, while self-hosted customers can request access.
The business case is narrow but important. Deepgram reports 91.59% drug-name recognition in batch and 91.28% in streaming, making the model a targeted alternative to broader clinical transcription models for pharmacy IVRs, refill automation and healthcare voice agents.
Nova-3 Pharma narrows the model around medication language
Nova-3 Pharma is a specialized option in the Nova-3 family rather than a general speech model. Deepgram’s model documentation describes it as optimized for drug names, dosages and medication terminology. The company says it uses the same architecture and compute footprint as Nova-3 Medical.
That makes the operational change small for existing Deepgram users. Developers select `model=nova-3-pharma` and can keep the same batch or streaming transcription architecture. Hosted access is public. Self-hosted deployment requires a request through a Deepgram account representative.
The model supports English variants including the United States, United Kingdom, Australia, Canada, Ireland, India and New Zealand. Unlike general Nova-3, it is not a multilingual model.
Drug-name recognition is the headline metric, but the benchmark is vendor-run
Deepgram reports a 91.59% Keyword Recognition Rate for drug names in batch and 91.28% in streaming. It also reports Word Error Rates of 10.17% for batch and 11.69% for streaming, which it says are the best results among the models it evaluated.
| Vendor-reported metric | Batch | Streaming |
|---|---|---|
| Drug-name KRR | 91.59% | 91.28% |
| Overall WER | 10.17% | 11.69% |
| Access | Hosted API | Hosted API |
KRR measures whether the specific drug names spoken in the audio are captured correctly. WER measures errors across the whole transcript, where lower is better.
The figures are useful because they measure the failure mode the product targets. They are still Deepgram’s evaluation, and the public material does not provide enough independent replication to treat the ranking as settled. Buyers should test the medication list, accents and audio conditions found in their own calls.
Nova-3 Medical remains the more general clinical option. It covers broader symptoms, diagnoses, procedures and clinical documentation. Nova-3 Pharma is the better test when medication terminology is the dominant risk.
Pricing is less clear than access
Deepgram’s public pricing page lists Nova-3 general rates, including promotional streaming pricing of $0.0048 per minute for monolingual pay-as-you-go and $0.0043 per minute for pre-recorded audio. The page does not list a separate Nova-3 Pharma line item.
That means teams should not assume the specialist model uses the general Nova-3 rate. Deepgram does state that pay-as-you-go and Growth plans include access to public model endpoints, but regulated deployments, self-hosting and enterprise terms can change the commercial picture.
For healthcare buyers, deployment controls matter as much as transcription price. Deepgram says enterprise deployments can use HIPAA-eligible infrastructure and Zero Data Retention. Its pricing page says Business Associate Agreements are available for enterprise customers handling protected health information.
What pharmacy and healthcare voice teams should test first
Start with a set of real medication calls that already produced corrections, failed refills or agent escalations. Score drug-name recognition separately from overall WER, because a transcript can look accurate while still missing the medication that drives the workflow.
Teams already on Nova-3 Medical can run a direct A/B test with minimal integration work. Nova-3 Pharma is worth testing now for medication-heavy workflows. General clinical dictation teams can keep Nova-3 Medical unless drug-name errors are a measurable problem.
What this means
Pharmacy platforms, prescription-refill systems and healthcare voice-agent teams should test Nova-3 Pharma now. The first workload should be a medication-heavy call set with known transcription failures and human corrections.
For those buyers, a specialist model can justify switching if drug-name KRR improves without hurting overall transcript quality or latency. Broader clinical documentation teams can wait, because Nova-3 Medical covers a wider vocabulary and the new Pharma benchmark remains vendor-reported.
For more information, visit the official announcement of Nova-3 Pharma on the Deepgram website.
