ScribeCheck: verified medical transcription that flags its own uncertainty

The problem

Doctors dictate notes under time pressure, and a single misheard drug name or dose in a transcript can become a chart error. Generic transcription tools miss speaker attribution, mishandle medical vocabulary and accents, and file text with no signal about which passages are shaky. Small clinics cannot afford enterprise scribe suites or full-time human scribes, so notes are either late or unchecked.

The idea

Why now

NVIDIA's Nemotron 3.5 Transcribe, released 28 Oct 2026, gives small teams an open model that combines transcription, speaker attribution and translation in one checkpoint, plus fine-tuning recipes and a medical case study where fine-tuning cut word error rate 45.8% relative on held-out speakers. The Granary v2 pipeline supplies the data curation workflow for domain adaptation. A year ago this stack required stitching four separate models together.

What it combines

Three capabilities combine: Nemotron 3.5 Transcribe (speaker-attributed, promptable transcription that can be fine-tuned per domain), Jev-style typed judgments (a decision layer that scores each span's trustworthiness and routes only shaky spans to humans), and the Granary v2 curation pipeline (domain data preparation). The mix matters because transcription alone produces plausible errors, and gating alone has nothing worth gating; together they give a small clinic an auditable, self-checking scribe.

MVP

Weekend scope: take the open transcription model, fine-tune it on a small public medical speech sample, and build a simple review page that plays back audio beside highlighted low-confidence drug and dose spans for one test clinic's recorded consultations. Deliberately skip: EHR integration, real-time streaming, multi-clinic tenancy, and automated note formatting.

How it makes money

The moat

Data and workflow, not model weights: each clinic's fine-tuned model on its own doctors, accents and formulary improves with that clinic's corrections, and the review workflow embeds into the clinic's charting routine. A large incumbent selling a one-size-fits-all enterprise scribe would have to rebuild around per-clinic fine-tuning, on-premise deployment and per-hour verification pricing, which undercuts its per-seat subscription revenue.

Distribution

B2B via small and mid-size clinics, starting with independent practices in India where the reference benchmark used Indian doctors' recordings. Wedge: partner with clinic-management software vendors and regional doctor associations that already own the clinic relationship, offering the service as an add-on to their practice software rather than selling clinic by clinic.

Why it wins

Enterprise scribes sell closed, per-clinician subscriptions and keep audio and models in their cloud. Human scribe services are slow and cost several hundred dollars per clinician per month. ScribeCheck is self-hosted, per-clinic fine-tuned, and sells verification rather than raw transcription, so the buyer pays for checked notes, not for model output.

Risks

Biggest risk is a confident-sounding error slipping past the gate into a patient record. The MVP de-risks this by never auto-filing flagged spans and by measuring gate recall on a labelled set of drug and dose mentions before any clinic pilot.

Build it with

• NVIDIA Nemotron 3.5 Transcribe: open-weight speech model that transcribes, translates and attributes speakers in one passSpeaker-attributed, promptable transcription plus fine-tuning recipes and the medical case study.
• Jev (TypeSafe AI)Typed judgment layer for scoring span trustworthiness and gating uncertain passages to human review.

Repo to start from

scribecheck — open service that fine-tunes speaker-attributed transcription per clinic and routes low-confidence clinical spans to human review before filing.

Evidence

• NVIDIA Launches Open-Weight Speech Model and 463,062-Hour Audio Dataset
• NVIDIA Nemotron 3.5 Transcribe and Granary v2
• The Real ROI of AI Scribes for Small and Mid-Sized Clinics