"Eén letsel" and "geen letsel" sound nearly the same at dictation speed. On the page they mean opposite things.
That is the problem in one line. Radiologists speak quickly, in dense terminology, in Dutch and French, with words everyday speech never uses. Speech models are mostly built for everyday speech. And dictation comes first: every later step in a report works from the words it wrote down.
So we decided that hearing radiology should be something Revelar can do itself. We started from an open speech model and taught it this one job. Our standard dictation still runs on a commercial engine, and it stays there until our own model has proven itself.
What we measured
In a first small test, Dutch speakers read fictional radiology reports aloud. These recordings never trained the model and were never used to choose it. We scored six open speech models, unchanged, against ours.
From open model to beta
What those numbers do not say
A good number on a small test proves little, so here are the limits as we know them.
- The test is small. It shows a direction. We will not compare ourselves with any commercial engine until many more voices have been scored.
- Some voices are harder. On one founder's own dictations of fictional reports the same model gets about 10% of words wrong.
- Critical words still go wrong. Negations, sides, numbers and units: three errors in this test. That is why the beta tells you to check every word.
- Live is not yet measured. These scores come from whole recordings. The beta reads while you speak, and we have not yet scored that on real voices.
- French is not ready. Our best French model gets about 19% of words wrong, so the beta is Dutch only.
Where the remaining errors sit
Most of what the beta still gets wrong is radiology vocabulary itself. That tells us where the next gains have to come from.
Critical words: what we are investigating
A few words carry far more weight than the rest. One wrong sound can turn a finding around, and the sentence still reads perfectly well. So we track these words separately from the overall score.
Our research question is simple: can a second, separate listen point out the word a radiologist should look at again? This is research. None of it is in Revelar today, and the radiologist checks every word.
The design we are testing
- HearThe model writes the dictation.
- FindNegations, sides, numbers, units and change words are marked.
- Listen againA separate listener hears only that stretch of sound.
- Point outIf they differ, the word is shown to the radiologist. It is never changed.
Ask other listeners
Several models hear the same passage, and we look where they disagree. What we learned: they often share the same confident mistake, so disagreement alone misses too much.
Forced choice
A listener gets both readings, such as "één" or "geen", and picks what the sound supports. A small hint for negations and sides, not strong enough yet.
A truly independent listener
Both tests point the same way: the second listener has to hear differently from the first. That is what we are looking for now.
Where your voice goes
The beta runs on our own speech server in Belgium, on Google Cloud. The server stores and logs no audio and no text. Dictations in the product never train the model, and no patient recordings or hospital data were used to train it.
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 it
Open the trial workspace for up to 5 minutes of Dutch dictation a day, no account needed. Please use fictional cases. Text appears 1 to 2 seconds after you speak and becomes final when you pause.