Medical voice recognition software turns clinician speech into clinical text, either by direct dictation or by listening to a visit and drafting the note. Vendors compete on accuracy claims that almost none of them publish data for. This guide covers the categories, the measured error rates, and how to test the claims yourself.
One naming note before the comparison. Medical speech recognition software, medical dictation software and speech to text for medical notes all describe this same category, and vendors use the labels interchangeably. In speech engineering, voice recognition usually means working out who is speaking and speech recognition means working out what was said. Product marketing has collapsed that distinction, so read every specification sheet as describing transcription unless it says otherwise.
The three categories, and what each one is for
Buyers usually start by comparing products. Start by comparing categories, because the three groups fail in different ways and are bought by different people.
| Category | What it does | Typical buyer | Named examples |
|---|---|---|---|
| Front-end dictation | You speak, text appears in the EHR field as you talk, you correct it live | Individual clinician or department | Dragon Medical One, Philips SpeechLive, Invox Dictation, VoiceboxMD (which also sells an ambient mode) |
| Ambient scribes | Listens to the visit, drafts a structured note afterwards for review | Practice or health system | Abridge, Suki, Nabla, Freed |
| Transcription APIs | Raw speech to text, you build the product around it | Engineering team | Amazon Transcribe Medical, Deepgram |
Front-end dictation puts the clinician in the correction loop in real time. Errors are visible immediately, and the cost is attention during the encounter.
Ambient scribes move correction to after the visit. That buys back attention in the room, and it moves the review burden to a moment when the clinician no longer remembers exactly what was said.
Transcription APIs are components rather than products. Amazon Transcribe Medical states support for primary care and specialty areas including cardiology, neurology, obstetrics-gynecology, pediatrics, oncology, radiology and urology, and it offers both batch and real-time speech to text. It publishes no accuracy percentage for any of them.
That absence is the pattern rather than the exception. Across the vendor pages and the roundups ranking for this category, product accuracy is either described in adjectives or stated as a round percentage with no test corpus, error definition or noise condition attached. VoiceboxMD, named in the table above, advertises 99% accuracy on its own homepage without naming a test set. A percentage with no corpus and no scoring method behind it is a claim, not a measurement.
How medical voice recognition software works
Every product in the table runs the same first stage. Audio is captured, segmented into speech and silence by a voice activity detector, and passed to an acoustic model that produces candidate text. A language model then rescores those candidates using context, which is why "hypertension" wins over "hypertensive" in one sentence and loses in the next.
The medical part is mostly vocabulary and bias. General models are trained on general speech, so drug names, anatomical terms and dosage formats are rare tokens. Vendors address this by fine-tuning on clinical audio, by biasing the decoder towards a medical lexicon, or by adding a correction pass that maps near-misses onto known terms.
Ambient scribes add a second stage. After transcription, a large language model reorganises the conversation into a structured note. That stage is where content can be added, dropped or reworded in ways the transcript itself would not predict, because summarisation is a generative step rather than a transcription one.
The practical consequence is that an ambient scribe has two places to go wrong and a dictation tool has one. Evaluate them accordingly.
The error rates the category does not advertise
The most detailed published measurement of dictation accuracy in real clinical use that this guide found is a 2018 JAMA Network Open study by Zhou and colleagues. The researchers compared 217 notes from 144 physicians at two health care organisations against a criterion standard built from the original audio.
The finding, in their words: "The error rate in SR notes was 7.4% (ie, 7.4 errors per 100 words). It decreased to 0.4% after transcriptionist review and 0.3% in SNs." Signed notes are the version the physician approved.
Two things follow. Seven errors per hundred words is the raw output of speech recognition, not what reaches the chart. In that workflow a professional transcriptionist is the review layer, and the review layer removes most of it. If you buy a product that removes the transcriptionist, you have moved that work onto the clinician rather than eliminated it.
The same study counted how many notes contained at least one error, which is the number that matters if you are the patient in one of them: 96.3% of the speech recognition notes, and 42.4% of the notes physicians had already signed. A low average error rate and a high share of affected documents are not in tension. Errors are spread thin across many notes rather than concentrated in a few bad ones.
Read the caveats before you reuse those numbers. The study measured back-end dictation, in which the engine output is edited by a professional medical transcriptionist before the physician reviews and signs it, at two organisations in 2016, on Dragon Medical 360 | eScription. It did not measure front-end dictation, where the clinician corrects on screen during the encounter, and it did not measure ambient scribes or any current product version. It establishes the order of magnitude, which is more than any vendor page offers.
A failure mode that word error rate does not catch
Errors are not only mistranscriptions. Speech recognition built on large models can produce text with no acoustic basis at all.
A peer-reviewed study presented at the 2024 ACM Conference on Fairness, Accountability, and Transparency, Careless Whisper by Koenecke and colleagues, evaluated OpenAI's Whisper as of 2023 and found that roughly 1% of audio transcriptions contained entire hallucinated phrases or sentences that did not exist in any form in the underlying audio. Of those hallucinations, 38% carried explicit harms, including perpetuating violence and implying false authority.
Carry the caveat with the number. That measurement was made on AphasiaBank recordings, a corpus of speakers with and without aphasia, not on clinical dictation, and it describes one model at one point in time. On a subset of 6,046 segments matched on demographics, hallucination rates were 1.8% for speakers with aphasia against 1.1% for the control group, and the study links them to longer non-vocal durations, which is to say pauses.
Pauses are not rare in a clinical encounter. Neither is disordered speech. A metric that counts substitutions against a reference transcript will score an invented sentence as a handful of insertions, which understates what it is.
What HIPAA actually asks of a vendor
Compliance pages compress this into a logo. The obligation is specific, and it sits with you rather than with the vendor.
Under 45 CFR 164.308(b)(1), a covered entity "may permit a business associate to create, receive, maintain, or transmit electronic protected health information on the covered entity's behalf only if the covered entity obtains satisfactory assurances, in accordance with § 164.314(a), that the business associate will appropriately safeguard the information."
The operative words are "obtains satisfactory assurances". A signed business associate agreement is how that is normally evidenced. No product setting substitutes for it.
Watch the vendor's exact wording. Amazon describes Transcribe Medical as "HIPAA-eligible", which is a statement about what the service can be used for under an agreement, not a declaration that your deployment is compliant. Vendors using "HIPAA-compliant" as a bare adjective are describing a posture, not a contractual position. Ask which entity signs, and what the agreement covers.
How to test medical voice recognition software before you buy
Vendor accuracy claims are unfalsifiable as written, so the only useful number is the one you produce on your own vocabulary. This is a bounded exercise rather than a procurement project, and most of the cost is the hand transcription, so budget for that before you start.
Build a reference set from your own audio. Collect 30 to 50 real encounters across your actual specialty mix, accents and rooms. Transcribe them by hand. That hand transcript is the only ground truth you will get.
Score entities, not just words. Overall word error rate is dominated by common words. Drug names, dosages, laterality, negation and numbers are what change care. Count those separately: a system at 4% overall that drops "no" before "evidence of" is worse than one at 7% that never touches negation.
Repeat the same input. A system that transcribes a recording correctly once has not been shown to do it repeatably. Cekura's own voice agent benchmark is built on that principle: a frozen study of 8 configurations across 82 scenarios with 3 retained repeats each, ranked on pass cubed, the share of scenarios passing on all three runs. Providers chose their own models and speech components; what was held constant is the prompt, the tool definitions and the test data. That benchmark covers voice agent platforms rather than clinical dictation, so its figures do not transfer to this decision. The method does.
Test in the room, not in the quiet. Record with the door open, the suction running and two people talking. Clean-audio accuracy is the number vendors demo.
Decide who reviews, and price it. The JAMA figures show review doing most of the accuracy work. If a product's pitch is that review is no longer needed, that is the claim to test hardest.
Cekura applies the same protocol to conversational systems, running scripted and generated scenarios against an agent, scoring each run, and repeating runs to separate a genuine pass from a lucky one.
Where this meets patient-facing voice agents
Dictation and ambient scribes point a microphone at a clinician. Some healthcare speech work points it at a patient instead, in intake, scheduling, refill requests and discharge follow-up. The recognition problem is harder there: unpredictable audio, no clinician correcting in real time, and an automated action at the end.
Cekura tests, monitors and self-improves those voice and chat agents. Cekura generates scenario coverage across accents and edge cases, scores each transcript against metrics you define, so a dropped negation or a wrong dosage counts as a failure rather than as one more word error, and monitors production calls so accuracy regressions surface as alerts rather than as complaints. Our guide to conversational AI in healthcare covers the architecture decisions behind that, and our note on ASR accuracy testing covers why word error rate alone is not the whole picture, especially across languages.
If you are evaluating speech in a patient-facing workflow rather than a documentation one, Cekura can run that evaluation against your own scenarios.
What to ask before you sign
Most of these questions have a documented answer. A vendor that cannot produce one has told you something useful.
- Who signs the business associate agreement, and what does it cover? Ask whether subcontractors, including any model provider behind the product, are covered.
- Is audio retained, and for how long? Ask separately about audio, transcripts and the drafted note, because retention often differs across the three.
- Is our data used to train your models? Ask for the answer in the contract rather than on the marketing page.
- What accuracy figure do you publish, on what corpus? A number without a corpus is not a number. If the answer is a customer-reported satisfaction score, treat it as one.
- How does the product behave on unclear audio? Ask whether it flags low-confidence spans or silently produces fluent text. Fluent text with no flag is the harder condition to catch in review.
- What happens at renewal if accuracy drops? Model updates ship continuously, and the version you tested is not the version you will run in six months.
That last point is the one buyers skip. Acceptance testing at purchase is a snapshot, and the thing being purchased moves. Cekura re-runs a fixed scenario set on a schedule for exactly this reason, so a regression shows up as a failed run rather than as a complaint from a clinician three months later.
Frequently asked questions
Is medical voice recognition software accurate enough for clinical notes?
With a review step, yes, in the settings that have been measured. The 2018 JAMA Network Open study found 7.4 errors per 100 words in raw speech recognition output, falling to 0.3% in the physician-signed note. Those figures come from back-end dictation at two organisations in 2016, with a transcriptionist editing before sign-off, and do not describe front-end dictation, ambient scribes or any current product.
What is the difference between front-end dictation and an ambient scribe?
Front-end dictation transcribes what you deliberately speak, and you correct it live. An ambient scribe listens to the whole encounter and drafts a note you review afterwards. Dictation costs attention during the visit; ambient scribes move that cost to review, when recall is weaker.
Does buying HIPAA-eligible software make my practice compliant?
No. Under 45 CFR 164.308(b)(1) the covered entity must obtain satisfactory assurances that the business associate will safeguard electronic protected health information, normally through a business associate agreement. "HIPAA-eligible" describes what a service may be used for under such an agreement, not the compliance of your deployment.
Can speech recognition invent content that was never spoken?
Yes. The Careless Whisper study at ACM FAccT 2024 found roughly 1% of transcriptions contained entire hallucinated phrases absent from the audio, with 38% of those carrying explicit harms. That measurement was on AphasiaBank recordings of speakers with and without aphasia, using one model as of 2023, not on clinical dictation.
How should I measure accuracy on my own medical vocabulary?
Hand-transcribe 30 to 50 of your own recordings, then score drug names, dosages, laterality, negation and numbers separately from overall word error rate. Repeat each recording to check the result is stable, and record in the room you actually work in rather than a quiet one.
