Deepgram has released Nova-3 Pharma, a speech-to-text model built specifically for pharmaceutical use cases such as pharmacy IVRs, prescription refill automation, contact center voice agents, and clinical documentation. Trained on more than 5,000 drug names, it claims the top rank in both drug name recognition (Keyword Recognition Rate) and overall transcription accuracy (Word Error Rate) across batch and streaming, beating general-purpose and medical speech-to-text competitors. Benchmarks show 91.59% drug name KRR in batch and 91.28% in streaming, with WER of 10.17% (batch) and 11.69% (streaming). It shares the same architecture and compute footprint as Nova-3 Medical, is available now via Deepgram's hosted API and self-hosted deployment, and supports eight English locale variants, invoked with model=nova-3-pharma.
Table of contents
High accuracy for pharmaceutical terminologyNova-3 Pharma is #1 in drug name recognitionNova-3 Pharma is #1 in overall transcription accuracyPower pharmaceutical applications with Nova-3 PharmaDeploy Nova-3 Pharma nowQuestions this post answers
What is Deepgram's Nova-3 Pharma model and how accurate is it at recognizing drug names?
Nova-3 Pharma is a speech-to-text model built specifically for pharmaceutical vocabulary, trained on more than 5,000 drug names. It achieves 91.59% drug name Keyword Recognition Rate in batch transcription and 91.28% in streaming, ranking first among evaluated general-purpose and medical speech-to-text models on real English medical audio. Developers building pharmacy voice tools can compare speech-to-text accuracy claims like these on daily.dev.
How do I access Deepgram's Nova-3 Pharma model for transcription?
Nova-3 Pharma is available now through Deepgram's hosted API and self-hosted deployments, supporting both batch and streaming transcription. Developers select it by specifying model=nova-3-pharma, and it supports eight English variants including en-US, en-GB, en-AU, en-CA, en-IE, en-IN, and en-NZ. Anyone integrating specialized transcription models can track new speech-to-text releases on daily.dev.
What is the word error rate of Deepgram Nova-3 Pharma compared to other speech-to-text models?
Nova-3 Pharma achieved the lowest Word Error Rate among all evaluated models, at 10.17% in batch transcription and 11.69% in streaming. It shares the same architecture and compute footprint as Nova-3 Medical, combining Deepgram's medical speech foundation with pharma-specific training. Teams evaluating speech recognition accuracy across models can follow benchmark updates on daily.dev.
Share this post