One word can reverse a finding

We’re building and testing a Dutch speech model for radiology. Here’s why, what our early tests show, and what still needs work.

Lucas Van Houtven, co-founder. November 2026

Radiology reports start with the words we dictate. When speech recognition gets one of those words wrong, the error can make its way into the report unless we catch it.

Some mistakes are easy to spot. Others are not. A missing “no”, a wrong side or a number that looks plausible can change the meaning of a finding.

Nearly the same sound

één letselone lesion
geen letselno lesion
At dictation speed, these two are almost the same sound. On the page, they mean the opposite.

As a radiologist in training, I find that difficult to accept. We use dictation software every day, yet we often have to read back with the nagging question: did it actually hear me correctly?

I want dictation to be something we can feel more confident using, with fewer mistakes to catch and less worry about what might have slipped through. We are not there yet, and radiologists still need to check their reports. That is the problem we want to work on.

That’s why we started working on a speech model for radiology. We began with an open model and adapted it for Dutch radiology dictation. It’s still early: our standard dictation uses a commercial engine, and our own model is available as a beta while we continue testing it.

How the model performed

In a small early test, Dutch speakers read fictional radiology reports aloud. We scored six open speech models, unchanged, against ours. The result is encouraging, but the sample is small and these recordings informed some development decisions. We need more voices that were not part of that process before drawing stronger conclusions.

Open model27.9%
Open model31.4%
Open model31.4%
Open model32.2%
Open model32.7%
Open model33.7%
Revelar (beta)5.0%
Word error rate, lower is better. Same recordings for every model, scored whole. The open models were not trained on radiology. This chart does not cover commercial engines.

What that means in 100 words

28 wrong words in every 100

Best open model, unchanged

5 wrong words in every 100

Our beta

Each dot is one word. Rounded from 27.9% and 5.0% on the same recordings.

From open model to beta

27.9%Open base, unchanged
7.2%Early radiology version
5.5%Later version
5.0%Beta today
Word error rate on the same recordings. The step from the later version to the beta is within this test's margin of error.

What we can and can’t conclude

This is a starting point, not a measure of how the model will perform for every radiologist.

  • Voices differ. On one founder's dictations of fictional reports, the same model had a word error rate of about 10%.
  • Important words can still be wrong. This test included three errors on critical words. Every report needs to be checked.
  • We haven't scored live dictation yet. The results above come from complete recordings, not the live experience in the beta.
  • The beta is Dutch only. We are also working on French, but it is not ready for the beta.

What it still gets wrong

Most of the errors in this test were radiology terms. That gives us something to investigate, but the sample is too small to assume the same pattern will hold for other voices.

80% radiology terms10% numbers7% small words3% other

The mistakes that matter most

A word error rate treats every word alike. In a report, a wrong negation, side, number or unit can matter much more than a misspelled term. We track these separately so they don’t get lost in the overall score.

Negationone lesionno lesion
Sideleftright
Number5 mm15 mm
Unit5 mm5 cm
Changeincreaseddecreased
Illustrative examples, not necessarily errors observed in the test.

Why these errors are hard to see

Dictated

Eén letsel in de rechter long.

On the page

Geen letsel in de rechter long.

A made-up sentence. Both versions read perfectly well, and only one is what was said.

We’re investigating whether a second model listening to part of the recording could flag a word for the radiologist to check. It would suggest a review, not change the report.

The design we are testing

  1. HearThe model writes the dictation.
  2. FindNegations, sides, numbers, units and change words are marked.
  3. Listen againA second model hears only that stretch of sound.
  4. Point outIf they differ, the word is shown to the radiologist. It is never changed.
Research design. Not in the product.

Why a second listener is not enough

What we seeTwo models often make the same mistake, so neither flags it.
What we needA listener that goes wrong in different places, so disagreement means something.
Schematic, not measured data. Each circle stands for one model's mistakes.
Tested

Ask other models

Several models hear the same passage, and we look where they disagree. They often share the same confident mistake.

Tested

Forced choice

A model gets both readings, such as "één" or "geen", and picks what the sound supports. Not reliable enough yet.

Open

A truly independent listener

The second listener has to hear differently from the first. That is what we are looking for now.

So far, this has been harder than it sounds. Different models often make the same mistake, and asking a model to choose between two possible words has not been reliable enough. This is research, not a feature in Revelar.

How the beta handles your dictation

The beta runs on our speech server in Belgium, hosted on Google Cloud. The server does not store or log audio or text. Dictations made in the product are not used to train the model, and no patient recordings or hospital data were used to train it.

Your browserSends the sound over an encrypted connection
Our server in BelgiumTurns sound into text. Keeps no audio and no text
Your reportThe text comes back to you, to check and sign

This covers the speech model. Other AI help in the workspace has its own processing, described on our Trust page. The beta is not used on hospital deployments.

Try the beta

You can try up to 5 minutes of Dutch dictation a day in the trial workspace, without an account. Please use fictional cases. Text appears 1 to 2 seconds after you speak and becomes final when you pause.