MayScribe whitepaper preview

Overview

The Case for Keeping Dictation Inside Your Own Walls

Hospitals have spent two decades buying dictation software that lives in someone else's cloud. The costs are now well documented: patient audio accumulates in vendor data centers, transcription errors cluster around the exact things that hurt patients (drug names, doses, units), and when the vendor has an outage or breach, every hospital on the platform goes down with it.

MayScribe takes a different position. Clinical dictation is infrastructure, and infrastructure that touches protected health information belongs inside the hospital's own security boundary. MayScribe is a self-hosted, real-time medical dictation system that runs entirely within the customer's cloud environment, on GPU instances the customer controls. Audio is processed in memory and discarded when the session ends. Nothing is retained.

"Between the speech model and the medical record sits a deterministic verification layer that checks doses, units, laterality, and sound-alike drug names against curated clinical vocabularies before a single word is committed to the note."

The Problem

What Goes Wrong with Clinical Dictation Today

96.3%

of unedited dictated notes contain at least one error

JAMA Network Open, 2018

30,000+

clinicians exposed to AI hallucination via Whisper-based tools

AP, Oct 2024

$7.42M

average healthcare data breach cost — highest of any industry, 14 years running

IBM, 2025

Errors Are Frequent, and the Dangerous Ones Look Ordinary

A 2018 JAMA Network Open study of 217 notes found 7.4 errors per 100 words, with 96.3% of unedited notes containing at least one error. About 1 in 6 errors involved clinical information. The Joint Commission documented cases like a spoken order for 40 mg of Lasix captured as 400 mg.

Newer Models Fabricate Fluent Text

In October 2024, an AP investigation found OpenAI's Whisper model, used by 30,000+ clinicians across ~40 health systems, fabricates text never spoken, including invented medical treatments. Academic research found hallucinated content clustering around pauses common in clinical conversation.

Cloud Concentration Creates Systemic Risk

The 2017 NotPetya attack took down Nuance Communications for weeks — $68M in lost revenue and $24M in remediation. IBM's 2025 report puts the average healthcare breach at $7.42M, the highest of any industry for 14 consecutive years.

Security Architecture

Security by Architecture, Not by Policy

Most vendor security stories are stacks of policy: we promise not to look, we promise to delete. MayScribe's position is that the strongest controls are structural — the system is built so the risky thing cannot happen, rather than promised not to.

Deployed Inside Your Own Cloud

Your VPC. Your subnets. No internet gateway in the audio path. Traffic moves over site-to-site VPN with mutual TLS. There is no MayScribe cloud in the path — PHI never transits infrastructure you don't control.

Zero Audio Retention

Audio streams into memory-backed buffers, is transcribed, and discarded when the session closes. No persistent storage. No retention setting to configure. An attacker who fully compromises the environment finds no archive of patient voice recordings, because none exists.

Tamper-Proof Audit Trail

Every session produces an append-only audit record: who dictated, when, into which system, what verification checks ran, which spans were flagged, and what the clinician did. Written with write-once object locking; cannot be silently altered, even by an administrator.

Hash-Pinned Model Integrity

Speech and verification models are self-hosted with hash-pinned weights. No external network calls at inference time. The artifacts running in production are cryptographically the artifacts that were reviewed and approved.

Threat Model

What this architecture removes.

Threat	How MayScribe eliminates it
Vendor cloud compromise	No vendor cloud exists in the data path
Stored patient audio theft	Zero audio retention — nothing to steal
Transport interception	Mutual TLS inside a private VPN
Silent record tampering	Write-once audit storage with object locking
Model supply-chain drift	Hash-pinned weights, cryptographically verified
Per-seat cost scaling	Infrastructure model — marginal cost of next user is zero

Compliance

A Compliance Posture Your Team Can Verify

HIPAA Alignment Today

MayScribe maps cleanly onto HIPAA's Security Rule because the architecture was designed with its categories in mind. Technical safeguards: mutual TLS over a private VPN, no persistent audio at rest, role-based access, and unique authentication for every user and service. Administrative safeguards: documented risk analysis, access review, incident response, and workforce policies — available as living documents, not summaries on a trust page. BAA chain is complete from clinician to hardware.

SOC 2 Roadmap

No vendor can hand you a legitimate SOC 2 report on day one. MayScribe's controls are designed, documented, and operating now, mapped to the Trust Services Criteria. An independent SOC 2 examination is targeted within the first year of production deployment. In the interim, MayScribe provides the control matrix, policies, architecture documentation, and evidence samples under NDA.

6 Questions to Ask Any Dictation Vendor

1Where does the audio go, physically, and who holds root on those machines?
2How long is audio retained, and who can change it?
3Is patient data used to train or tune models, and is that contractual?
4What happened to customers during your last significant outage or breach?
5Which checks run on drug names, doses, and units before text enters the chart?
6Can our security team inspect the running system — not a diagram of it?

MayScribe's answers: your VPC · zero retention · never · not applicable by architecture · deterministic checks on every commit · yes.

Performance

Real-Time Engineering Without Shortcuts

Clinicians abandon dictation tools that make them wait. The engineering problem is that the two things hospitals want — immediate text and verified text — pull in opposite directions. MayScribe resolves the tension with a two-pass design instead of a compromise.

~0.4s

Median latency from end of speech to committed text at cursor

Design target

2-Pass

Streaming draft + background rescoring model running in tandem

Architecture

0 Retention

Entire pipeline runs in memory on GPU hosts inside your VPC

Deployment

STEP 1

Clinician Speaks

Push-to-talk with voice activity detection. Microphone open only when deliberately activated.

STEP 2

Pass 1 · Streaming Draft

Fast speech model tuned for latency. Partial text appears at cursor within a fraction of a second.

STEP 3

Pass 2 · Rescoring

Larger background model corrects word boundaries, rare terminology, and numbers on completed utterances.

STEP 4

Verification Layer

Deterministic clinical checks: doses, units, laterality, negations validated against curated vocabularies.

STEP 5

Commit or Hold

High-confidence spans commit instantly. Flagged spans held for single-keystroke clinician review.

STEP 6

Text at Cursor

Lands directly in the EHR — including virtualized desktop environments. No round trip to external services.

Clinical Accuracy

Accuracy Is a System Property, Not a Model Score

The errors that matter clinically are concentrated in a handful of token types, and the newest models add fluent fabrication to the list. A credible accuracy story has to be about the system around the model, not the model alone.

A Clinical Lexicon Underneath Everything

Verification runs against RxNorm, SNOMED CT, ICD-10-CM, and LOINC, together with the ISMP list of look-alike, sound-alike drug pairs. When a clinician says a drug name, the system resolves the utterance against a formulary-aware vocabulary — not pattern-matching syllables. Sound-alike pairs trigger review rather than silent commitment.

Deterministic Checks Between Model and Chart

Before any text is committed, a rule layer validates doses against plausible ranges, normalizes units deterministically (a spoken milligram cannot silently become a microgram), flags unsupported laterality terms, and tracks negations so a dropped word cannot reverse a clinical statement. These checks are rules, not model opinions. They behave the same way every time.

Human Judgment Spent Where Risk Actually Is

Clean, high-confidence dictation commits automatically. Roughly one span in every few dozen — doses, units, sided terms, negations the system is unsure about — is held for the clinician with alternatives ready for single-keystroke resolution. The machine handles the volume; human attention is reserved for the spans that carry real risk.

Error Categories MayScribe Targets

Relative clinical risk score (0–100).

Dose errors
95
Unit errors (mg vs mcg)
90
Negation drops
88
Laterality errors (left/right)
85
Sound-alike drug names
82
Word boundary errors
40
General vocabulary errors
25

High-risk error categories — doses, units, negations, laterality, and sound-alike drug names — all score above 80. MayScribe's deterministic rule layer and clinical lexicon are purpose-built to intercept precisely these categories before any text reaches the chart.

Pricing & Scale

Scaling on the Hospital's Terms

Cloud dictation is priced per seat per month, meaning the cost scales with exactly the thing a hospital wants to grow: the number of clinicians using it. Self-hosting inverts the model.

Infrastructure Pricing, Not Per-Seat Licensing

MayScribe runs as containerized inference services on a small number of GPU instances inside the hospital's cloud account. The design target for a community hospital's full dictation load is two GPU nodes, with capacity added by adding nodes, not by renegotiating licenses. The hospital pays its cloud provider for compute it controls, and the marginal cost of the next physician who starts dictating is effectively zero.

Tuned to Your Hospital, Not a Global Average

Because the system is deployed per customer, tuning is per customer too: the formulary in the lexicon is the hospital's formulary, the specialty vocabulary reflects the hospital's case mix, and performance is measured against that hospital's real acoustic conditions.

Availability That Doesn't Depend on Anyone Else

A hospital running MayScribe is not sharing fate with a thousand other customers on a vendor's platform. Its dictation capacity is its own infrastructure, inside its own disaster recovery posture. An incident elsewhere in the world does not reach it. Updates ship as versioned, hash-pinned artifacts the hospital applies on its own schedule, through the same change-management process it uses for any clinical system.

Dimension	Cloud dictation (typical)	MayScribe (self-hosted)
Pricing model	Per seat / per month	Infrastructure compute (you control)
Cost as you scale	Increases with every new clinician	Marginal cost of next user: ~$0
Audio data location	Vendor data centers	Your VPC only
Outage risk	Shared across all vendor customers	Isolated to your own environment
Update schedule	Vendor-controlled	Your change management process
Tuning	Global average	Per-hospital formulary & case mix

Built to Be Audited. Designed to Be Trusted.

Hospitals evaluating dictation vendors do not need to take any of this on faith. The architecture is inspectable in your own environment, and the roadmap to independent attestation is stated in plain terms. That is what it looks like when a vendor expects to be audited — and builds accordingly.

Security Whitepaper

Available under NDA to hospital security and compliance teams.

Control Matrix & Architecture Review

Full documentation provided to your risk team before contract.

Pilot Validation

Design-target figures validated per deployment during pilot.

MayScribe provides a security whitepaper, control matrix, and architecture review under NDA to hospital security and compliance teams. Design-target figures cited in this document are validated per deployment during pilot.

Clinical Dictation, Without the Vendor Risk