Gemini 3.5 Transcribe for clinical dictation: worth trialing, not deploying

What it is

Google's new speech-to-text model, delivered through the Gemini API. It provides low-latency, accurate transcription with utterance-based language detection, speaker diarization, word-level timestamps, Smart transcription, and custom vocabulary speech biasing. Gemini 3.5 Transcribe is a speech-to-text model based on Gemini's audio understanding capabilities. Two variants: a file-based model for pre-recorded audio and a streaming variant for live capture.

The next-work-session test

Concrete scenario: an internist finishes a 12-minute new-patient visit and wants a structured note in the EHR before the next appointment.

What changes with this model versus a generic ASR: Smart transcription seamlessly handles self-corrections (like "let's meet Tuesday—no, Wednesday"), removes filler words ("ums" and "ahs"), auto-formats output, and custom vocabulary lets you bias the recognizer toward drug names, procedure codes, and local specialist names you supply in the request. That's the part clinicians actually care about — fewer "metformin" → "metforming" corrections in the HPI.

What doesn't change: it still hands you plain text. You (or a wrapper app) still have to structure it into SOAP sections and push it into Epic, Cerner, or athenahealth. Nothing in the DeepMind post or the API docs describes an EHR connector.

Pricing

Public preview pricing, per third-party analysis of Google's own token math: approximately $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). Google reports a 2.6% word error rate on non-streaming English as measured by Artificial Analysis. Treat both numbers as vendor-adjacent — the pricing is derived from token rates, not a published per-minute SKU, and WER on clinical audio (accents, background noise, jargon-dense dictation) will be worse than the clean-English benchmark.

What we'd actually use it for

Ambient scribing pilots and back-office transcription — visit recordings fed into a documentation assistant, call-center QA on patient-services lines, transcribing recorded case conferences for the M&M archive. The realistic wedge is "cheaper, cleaner raw transcript than your current vendor" — not "AI clinical documentation."

Also useful: quick internal work like transcribing recorded interviews with clinicians during a workflow-redesign project, or captioning training videos for nursing staff. Non-PHI use cases where the price and the filler-word cleanup earn their keep on day one.

Limits

Read this section twice if you handle PHI.

HIPAA is not automatic. Gemini.google.com and the Gemini mobile app are consumer products. There is no BAA path for consumer Gemini. A BAA is only available through specific Google Cloud services under contract. You need to confirm in writing that the specific Gemini API endpoint serving gemini-3.5-transcribe is covered by your BAA and configured with the right data-residency, logging, and no-training controls before any patient audio touches it. Do not assume.

No native EHR integration. No FHIR out-of-the-box, no Epic App Orchard listing on launch, no HL7 v2 emitter. You're building the pipe.

"Smart" edits are a clinical risk. Auto-removal of self-corrections is great for meeting notes and dangerous for medication dosing ("give 10 — no, 20 milligrams"). Custom vocabulary biases the recognizer, but it does not guarantee the right choice between sound-alikes (hydralazine vs. hydroxyzine). Human review of every clinically actionable line remains mandatory.

Preview-tier stability. Public preview means SLAs, rate limits, and even model behavior can change. Not the surface you want under a production dictation workflow yet.

Long encounters. The file model handles audio up to about one hour per request; longer visits or full-day surgical recordings need chunking, which complicates diarization continuity.

Try it if

• You're evaluating ambient-scribe vendors and want a price/accuracy floor to benchmark them against.
• You have a non-PHI transcription workload (training, internal meetings, patient-education video captions) and are paying too much for it today.
• Your health-IT team already has a Google Cloud BAA and can stand up a controlled pilot with de-identified audio.
• You want the raw transcript layer under your own summarization prompts, not a black-box scribe product.

Skip it if

• You need a turnkey clinical documentation tool that lands notes in Epic tomorrow. Buy a scribe product (Abridge, Nuance DAX, Suki) instead.
• You cannot get a signed BAA covering this specific endpoint. Non-negotiable.
• Your workflow depends on preserving every verbal self-correction verbatim for medico-legal reasons — the "smart" cleanup works against you.
• You're a solo practice without engineering support. There is no clinician-facing UI here; it's an API.

Source: DeepMind's announcement post — Intelligent transcription with Gemini 3.5 Transcribe. Model behavior and options: the Gemini 3.5 Transcribe API 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.