Deepgram nova-3-pharma vs Sume STT: drug-name transcripts

Sume has no pharma model to pick. Speech-to-text runs on one public model id, sume/stt-1.0, and the request accepts no model selector, so you cannot ask for drug-name vocabulary the way Deepgram's new nova-3-pharma does. What Sume returns is words[] with timings, which lets a reviewer check each medication name against the audio.

Deepgram's side is from its changelog; Sume's side is from the API reference. Both read 2026-10-01.

What did Deepgram release?

The Deepgram changelog entry dated Sep 17, 2026 describes nova-3-pharma as a new Nova-3 model built for pharmaceutical vocabulary, focused on accurate drug-name recognition, for pharmacy and healthcare voice-agent workflows. It is available in English for both batch and streaming.

What does Sume STT let me choose?

Little, by design. The request takes a public audio_url, an optional language_code, and an optional duration_seconds. The schema says provider knobs such as diarize and tag_audio_events are fixed server-side, and provider model ids stay internal.

How do I verify a drug name with Sume?

Transcribe, then treat every medication token as unconfirmed. Use the start and end seconds on that word to cut a short clip and have a person listen to it. Word timings are a review aid, not a clinical check, and Sume's docs make no accuracy claim for pharmaceutical terms.

If a drug name is wrong, a language hint will not fix it, because the hint selects a language, not a vocabulary. Plan a human or lookup step after transcription.

What should I do next?

If you need a vocabulary-tuned medical model today, use the vendor's. If you stay with Sume, keep the job id and word timings with each transcript so reviews are repeatable. For a fuller account of what Sume STT returns for medical audio, see what Sume STT does for medical transcription.