A doctor in Sarajevo says “krvni pritisak”. A doctor in Zagreb says “krvni tlak”. A doctor in Belgrade often says both, and a patient from Banja Luka says “tenzija”. All of these mean the same thing, and in a structured report they have to end up as one consistent term. For an AI tool meant to support doctors in Bosnia, Croatia and Serbia, the language is a harder task than the medicine. This article explains why, and how to tell a tool that works from one that only looks like it does.
Why the language is hard
A common assumption is that a tool built for English only needs localization for other languages. In a medical setting that assumption fails in several places.
- Little medical text. There is far less medical text in Bosnian, Croatian and Serbian than in English, and tools are trained mostly on English.
- Several standards. Ijekavian and ekavian, Latin and Cyrillic script, regional vocabulary: “sedmica” and “tjedan”, “voz” and “vlak”.
- Latin in everyday speech. Doctors dictate Latin diagnoses as a matter of course. A tool must treat them as proper medical terms, not as errors.
- Free word order. The languages allow many more word-order permutations than English, which makes sentence structure harder to follow.
Many words for one thing
The clearest example is medical synonyms. The doctor and the patient use different words for the same finding, and the report needs one.
- High blood pressure: krvni tlak, krvni pritisak, tenzija, visok pritisak
- X-ray: rendgen, RTG, snimak
- Kidney stone: kamen u bubregu, kalkulus, nefrolitijaza
- Dizziness: vrtoglavica, vertigo, omaglica, “glava me okreće”
- Cough: kašalj, tussis, suh kašalj
A patient describes the history the way they speak. A doctor wants the report in the wording medicine expects. A tool that does not bridge the two leaves three different terms for one thing, and the doctor corrects them by hand every time. Whatever the tool does, the wording of the report follows the doctor: the doctor decides which term stays, and reviews the report before signing.
Latin in the doctor’s voice
A typical dictation mixes the local language with medical Latin: “Diabetes mellitus typi 2 s dobrom regulacijom”, “status post cholecystectomiam laparoscopicam, bez komplikacija”. This is not a mistake. It is the professional language taught at medical school. A generic speech tool trained on English sees “a local language with unknown words”. It then either tries to “correct” the Latin into something unreadable, treats every Latin word as uncertain, or makes wrong substitutions, so “typi 2” becomes “tipi dva”.
Three languages in one sentence
A realistic dictation from an internal medicine clinic: “Pacijent ima history hipertenzije deset godina, sad treba dodati diuretik jer se javlja peripheral edema.” Local grammar, Latin terms and English technical words in a single sentence, without a pause. A tool that detects only one language ignores the rest. A tool that translates everything turns “peripheral edema” into free text instead of a clinical finding. For regional medicine, mixed language has to be the normal case, not the exception.
Dialects and pace
A doctor from Sarajevo, one from Mostar, one from Banja Luka and one from Niš do not speak identically, even when they say the same thing. Vocabulary, accent, speed and the pace of dictation all differ. The only real answer is a tool built and tested with regionally varied medical speech.
What separates a tool that works
- Medical vocabulary taken seriously. Latin terms, English terms used in medicine and local variants are treated as expected words, not rare ones.
- Post-processing that knows medical language. After the speech is recognized, a language model arranges it into the sections of your template, with consistent terms and punctuation.
- You stay in control. You read and edit the report, and the doctor signs. The tool writes what you said and does not decide on the diagnosis.
AiMED Dictate is tuned for Bosnian, Croatian and Serbian medical speech, plus English and German, and you can dictate in more than 30 languages.
How to test it in practice
- Dictate five of your most common reports, at your natural pace and with your natural mixing of languages.
- Count the corrections in each transcribed report.
- Measure the time from the end of dictation to a report ready for export.
- Compare it with typing, which is the current standard.
If a report needs minutes of corrections, the tool probably has no real grounding in regional medical language. If it needs a few seconds, it probably does. You can do this in one afternoon with any tool that lets you try it. For the wider picture, read AI tools for doctors in Bosnia, Croatia and Serbia, or get in touch and we will show you AiMED Dictate on a report from your specialty.
