Medical dictation software: Dragon, ambient scribes and when to build
Short answer: dictation and ambient scribing solve different problems and the distinction decides your purchase. Dictation transcribes what you deliberately narrate. An ambient scribe listens to the consultation you were having anyway. If you are still narrating, you are doing a second pass over work you already did — and that second pass is the thing worth removing.
Dragon Medical One is the incumbent and it is genuinely good at what it does. Whether what it does is what you need is the question worth thinking about before signing a three-year term.
What each category actually is
| Dictation (ASR) | Ambient scribe | |
|---|---|---|
| Input | You narrate, deliberately | The consultation itself |
| When | After the encounter | During, passively |
| Output | Text you dictated | Draft note or filled forms |
| Reported error rate | 7–11% | 1–3% |
| Clinician behaviour | Changes — you compose aloud | Unchanged |
| Best at | Precise control of wording | Removing the second pass |
That error-rate gap is counterintuitive and worth understanding. Dictation is transcribing a single clear speaker in controlled conditions, which should be easier. Ambient capture handles overlapping speakers and clinic noise, which is harder — but modern ambient systems are measured on the structured output after downstream processing and correction, while dictation ASR is measured on raw transcription. Different denominators, not a straight capability comparison.
What Dragon Medical One costs
Reported pricing, since Microsoft does not publish a list:
- $79–99 per user per month, billed annually — roughly $99 on a one-year term, $89 on two years, $79 on three.
- ~$525 one-time implementation per new user for install and training.
- $300–650 for a microphone (Philips SpeechMike, Nuance PowerMic III). Not required, but heavy users buy one.
The $79 rate is tied to the three-year commitment, and early cancellation generally still owes the remaining contract value. Price the three-year total, not the monthly headline: at 25 clinicians on the three-year rate that is roughly $71,000 in licences plus $13,000 implementation before hardware.
It is the strongest pure dictation tool available, with deep EHR integration and a claimed 99% accuracy across a reported 550,000+ users. If dictation is what you want, this is the default and you should probably just buy it.
Cost per clinician per year is the number that eventually pushes larger practices toward building their own.
The question that decides it
Not accuracy. Not price. What does your chart actually require?
If your output is prose — a narrative note you compose in your own words — dictation is the right tool and Dragon is the best of them. Precise control over wording is a real feature, and ambient systems give you less of it.
If your output is a set of structured mandatory forms, dictation barely helps. You still decide which fact belongs in which field of which form, and you still key it in. The transcription was never the bottleneck. We wrote about this at length in what is an AI medical scribe and it remains the most common scoping error we see.
Count it honestly: after your current tool runs, how many fields does a clinician still complete by hand? That number is the size of the problem, and dictation software does not reduce it.
Where ambient sits
Ambient clinical documentation removes the second pass entirely. Nothing about the clinician’s behaviour changes — no narration register, no wake word, no pausing to dictate. Adoption is markedly better than dictation tools for exactly this reason.
The trade is control. You get a draft you review rather than words you chose. For most clinicians reviewing a good draft is faster than composing aloud, but not for everyone and not in every specialty, and it is worth piloting rather than assuming.
Cost is the other trade: ambient products run substantially above dictation pricing. We put real numbers on that comparison in what AI medical scribes actually cost.
The categories buyers are actually comparing
Shop for this and you will meet four labels for overlapping things, which is most of why the market is confusing.
Medical dictation software is the desktop and browser class: you speak, it types into the field your cursor is in. Dragon is the reference point and the thing everything else is measured against.
A medical dictation app is the same capability on a phone, usually recording to a queue that syncs rather than typing live. Clinicians moving between rooms tend to prefer this, and the best medical dictation app for a given practice is usually whichever one their EHR vendor already integrates, because the alternative is copy and paste.
Medical voice recognition software is the same thing described by its mechanism rather than its use. When a vendor sells voice recognition in healthcare as a distinct product category, ask what it does that dictation does not; the answer is frequently nothing.
Medical transcription software historically meant routing audio to a human typist, and now usually means automated speech recognition with a review step. AI medical transcription software is the current label for the same job with a better acoustic model behind it.
None of these listen to a consultation. All of them transcribe deliberate narration, which is the distinction that actually matters and the one the category names obscure.
Where accuracy actually breaks for physician dictation
Physician dictation software is judged on word error rate, and the aggregate figure is close to useless because errors are not evenly distributed.
Clean narration in a quiet room is close to solved. What breaks is drug names that sound like other drug names, dosages where a decimal point changes the meaning, abbreviations that expand differently by specialty, and accented speech that the acoustic model saw less of in training.
That last one is worth stating plainly: dictation for doctors performs measurably worse for clinicians whose first language is not English, and vendors do not publish the breakdown. If a meaningful share of your medical staff is in that group, the aggregate accuracy figure in the proposal does not describe them.
The practical test is to record twenty real medical dictations from your own clinicians, including the ones with accents and the ones who speak quickly, and measure errors by category rather than in total. A system that is 97% accurate overall and wrong on dosages is not usable.
Hardware, and whether it still matters
Medical dictation devices were once the whole product: a handheld recorder with programmable buttons, docked to upload. Dedicated medical dictation hardware still sells, and for a clinician who dictates for hours a day the button placement and microphone quality genuinely justify it.
For everyone else the phone is the device, and the microphone in a recent phone is good enough that upgrading it is rarely where accuracy comes from. Spend the budget on the model and the integration instead.
When building beats both
We build custom medical scribes and we tell most people to buy one. Building is right in four situations, none of which are about price at small scale:
Your burden is forms, not a note. Products optimise for a good clinical note because that is the common denominator across their customers. If you file a pile of setting-specific structured forms, a beautiful note leaves most of the work in place. Our production build auto-fills more than ten mandatory forms from a single encounter, with no clinical fact entered twice — that de-duplication is where the hours come back.
Your EHR is not on the integration list. If you are on Epic, buy. If you are on something regional, older or homegrown, you will be quoted an integration project regardless, and at that point you are paying custom prices for something you do not control.
Data residency rules out the vendor. Audio, transcripts and generated notes are all PHI. If they cannot leave your boundary, most of the market is eliminated before pricing enters the conversation. See HIPAA-compliant AI architecture for what that actually requires.
Per-encounter pricing dominates your costs. Some contracts price per encounter rather than per clinician, which is worse for high-volume settings because your cost scales with the activity you are trying to make efficient.
How to evaluate, properly
Demos use clean audio and cooperative speakers. Insist on yours.
- Your accents, your specialty vocabulary, your room. Drug names and clinical terminology are exactly where general models fail and exactly where errors are consequential.
- Your forms. Not “can it write a note” — can it complete the documents you are required to file.
- Count what remains. Fields a clinician still fills by hand after the tool runs. That is the product.
- Feed it a difficult encounter. Does it flag uncertainty, or invent something plausible? A confident wrong field is worse than a blank one, because a blank field gets noticed.
- Price the full term. Licences plus implementation plus hardware plus the renewal rate, in writing.
Point four separates serious systems from impressive ones, and it is the one nobody tests in a demo.
The takeaway
Dictation software is a mature, well-solved category, and Dragon Medical One is the reasonable default if narrating is genuinely your workflow. The question worth asking first is whether narration is the burden at all. For a lot of providers the burden is structured forms, and no dictation tool — however accurate — touches that.
Sources
- HHS guidance on HIPAA and PHI
EpochC builds AI medical scribe development and clinical AI, including one auto-filling 10+ mandatory forms per consultation. Start a project — we will tell you honestly if a product solves it. Audio work runs on our the speech-to-text development behind it.