Gemini 3.5 Transcribe for clinical dictation: does it survive an EHR workflow?

What it is

Gemini 3.5 Transcribe is Google DeepMind's new speech-to-text model, built on the Gemini 3 audio stack. Gemini 3.5 Transcribe is a speech-to-text model based on Gemini's audio understanding capabilities. It provides low-latency, accurate transcription with utterance-based language detection, speaker diarization, word-level timestamps, Smart transcription features. It ships as two API endpoints — a batch/file model and a live streaming model — under the Gemini Developer API.

The next-work-session test

Concrete scenario: an internist finishes a 20-minute new-patient visit and needs a SOAP note in the EHR before the next patient arrives.

Today, most clinicians either dictate into Dragon Medical / Nuance DAX, type the note themselves, or use an ambient scribe like Abridge or Suki. Would Gemini 3.5 Transcribe change that next session? Not directly — it is a raw transcription API, not a clinical documentation product. There's no HL7/FHIR write-back, no SOAP structuring, no ICD-10 hinting, no EHR plug-in. What it could change is the workflow of a health-tech team building an internal scribe tool: they now have a cheaper, streaming-capable STT layer with speaker diarization and word timestamps to build against.

Pricing

Google publishes the rates on the Gemini API pricing page. Estimated pricing is based on 25 audio tokens per second for input and 175 text tokens per minute for output, for an effective blended rate of ~$0.009 per min for Live Transcribe. Third-party trackers put the file (batch) model at $0.005 per minute for the file model and $0.009 per minute for live, which undercuts Google's own Cloud Speech-to-Text v2 standard tier ($0.016/min). Treat those minute-based figures as derived estimates — the API bills in tokens, so real cost depends on audio density and output length. For HIPAA use, pricing is only half the question: consumer Gemini pricing does not carry a BAA. You need Vertex AI on Google Cloud.

What we'd actually use it for

Honestly? Two narrow things:

We would not put it in front of a clinician as-is for live patient encounters.

Limits

• Not HIPAA-compliant out of the box. Google Gemini: HIPAA compliance possible only with enterprise BAA, complex configuration, limited products and continuous oversight. Gemini can be used in HIPAA-regulated workflows through one specific route: Vertex AI on Google Cloud Platform, with a Business Associate Agreement (BAA) signed with Google, on a HIPAA-aligned Google Cloud account. Consumer Gemini at gemini.google.com is off-limits for PHI. Using the Gemini Developer API keys straight from AI Studio is the wrong door for a clinical workflow.
• No clinical fine-tuning claimed. Google's reported 2.6% word error rate on non-streaming English (as measured by Artificial Analysis) is on general audio, not on medical dictation with drug names, dosages, or anatomy. WER on clinical vocabulary is not published. Assume it's higher.
• No EHR integration. No Epic, Cerner, or Athena connector. No structured note output. No specialty templates. That's the job of a documentation product built on top of it.
• Preview status. Gemini 3.5 Transcribe entered public preview on August 26, 2026 — SLAs and versioning behavior will shift.
• Human review still required. For any PHI-adjacent output, a clinician has to read and sign. No STT model changes that.

Try it if

• You're a health-tech engineering team on Vertex AI with a signed BAA, evaluating STT layers for an ambient scribe or call-center product.
• You need cheap streaming transcription for non-PHI use cases (internal meetings, education, medical podcasts).
• You're benchmarking against Deepgram, AssemblyAI, or Cloud Speech-to-Text v2 and want a fresh data point on cost per minute and diarization quality.

Skip it if

• You're a clinician looking for a plug-and-play dictation replacement for Dragon or an ambient scribe like DAX, Abridge, or Suki. This is an API, not a product.
• Your organization has not signed a Google Cloud BAA, or your workflow routes through the consumer Gemini app or the plain Developer API.
• You need proven accuracy on medical terminology today — wait for a healthcare-specific benchmark or a vendor built on top of this model that publishes clinical WER.

Source: DeepMind's announcement post and the model documentation.

---

This content is for informational purposes only and is not medical advice. AI tools used with patient data must meet your organization's HIPAA and privacy requirements.