Scarce data
Little labeled Amharic clinical speech exists. We build consented datasets carefully, never scraping patient conversations without permission.
Bekur — AI & ScienceAddis Ababa · 2026
THE TECHNOLOGY — ቴክኖሎጂው
How BekurunderstandsAmharic & English.
Bekur listens to a real consultation — Amharic, English, and the code-switching in between — and writes the structured clinical note automatically. Here is how the system works, how we measure it, and how we keep patient data safe.
Bilingual speech recognition
Clinical SOAP structuring
Built for low-resource languages
FOLIO 01 · THE PIPELINE — ሂደቱ
From spoken words to a signed note.
A consultation is messy — two voices, two languages, interruptions, and clinical shorthand. Bekur turns that raw audio into a structured note in four stages, each one tuned for Ethiopian clinical reality.
01
Capture
መቅረጽBekur records the consultation through the clinician's device. Audio is processed securely and encrypted.
02
Recognize
ማወቅSpeech recognition transcribes Amharic and English in real time, tuned on Ethiopian accents, names, and the spoken patterns of a real clinic — not Western datasets.
03
Code-switch
ቋንቋ መቀየርDoctors mix Amharic and English mid-sentence. Our models follow the switch word by word, so a phrase that starts in Amharic and ends with an English drug name still resolves correctly.
04
Structure
ማዋቀርA clinical language model maps the conversation into a structured SOAP note — Subjective, Objective, Assessment, Plan — ready for the clinician to review, edit, and sign.
FOLIO 02 · ACCURACY & EVALUATION — ትክክለኛነት
Accuracy is earned, and measured.
A medical scribe is only useful if clinicians trust it. We treat accuracy as an engineering discipline — measured against clinician-reviewed ground truth, on real Ethiopian consultations, in both languages.
2 languagesAmharic + English
Recognized natively, including mid-sentence code-switching.
100%Clinician-reviewed
Every generated note is reviewed and signed by a clinician before it enters the record.
40%Of the workday
The share of a doctor's day documentation can consume — the time Bekur is built to give back.
Ground truth, not guesswork
We benchmark transcription and structuring against notes written by Ethiopian clinicians, so our metrics reflect the work as it is actually done.
Human in the loop
Bekur drafts; the clinician decides. Nothing is finalized without review, and every edit becomes a signal that makes the next note better.
Failure-aware by design
When the model is unsure — an unfamiliar drug, a noisy room — it flags the uncertainty rather than inventing detail. We optimize against confident mistakes.
FOLIO 03 · THE SCIENCE — ሳይንሱ
The science of low-resource languages
English medical AI rides on decades of digitized notes, open datasets, and billions of labeled words. Amharic has almost none of that. Building a clinical scribe for Ethiopia means doing the harder science: making models that learn from far less, in a language with rich morphology and a non-Latin script.
Amharic words inflect heavily — a single root can take dozens of forms — and clinical speech blends it with English drug names, lab terms, and abbreviations. Off-the-shelf models trained on Western data simply break here. So we build the data, the evaluation, and the models for the language as it is actually spoken in an Ethiopian consulting room.
This is the work: collecting and respecting consented clinical speech, tuning recognition for local accents and code-switching, and teaching a language model to structure messy bilingual dialogue into a note a doctor would sign. We build in the open because credible medical AI for low-resource languages should be earned in public.
“If technology only works in the languages with the most data, it will never reach the patients who need it most. We build for Amharic first, on purpose.”
Rich morphology
One Amharic root takes many forms. Recognition and structuring must handle inflection that Western tokenizers were never designed for.
Constant code-switching
Clinical Amharic is interleaved with English terms. Models must follow the switch fluidly rather than treating one language as noise.
Clinical correctness
A wrong drug or dose is unacceptable. We optimize against confident errors and keep a clinician in the loop on every note.
FOLIO 05 · FAQ — ጥያቄዎች
Questions, answered
What clinic and practice leaders ask before bringing Bekur to their team.
- What is Bekur?
- Bekur is an AI medical scribe built in Ethiopia. A doctor speaks naturally with a patient in Amharic or English, and Bekur writes a structured clinical note in seconds. It is the first medical scribe built to understand Amharic clinical conversation, including code-switching between Amharic and English.
- Does Bekur understand Amharic?
- Yes. Bekur is built specifically for Amharic clinical speech, including consultations where doctor and patient switch between Amharic and English. Most scribe tools are built for English only. Bekur is the first designed for the way medicine is actually spoken in Ethiopia.
- How much does Bekur cost?
- Bekur is charged per consultation, about the price of the visit itself. There are no subscriptions and no per-seat licences. The price includes the AI scribe, the structured SOAP note, and the patient record. Government and public programmes are priced by quote.
- Is Bekur free for doctors?
- Doctors who join the founding cohort use Bekur free during the pilot and after. The clinic or institution pays per consultation. Individual doctors on the waitlist get free early access while the product is being built.
- Who owns the patient record?
- The patient does. Each record is tied to the patient's phone number, follows them between clinics, and can be shared with any doctor in one tap. There is no app to install and no portal password to manage.
01
02
03
04
05
FOLIO 06 · SEE IT WORK
See the science in a real note.
Bekur listens in Amharic and English and writes the clinical note automatically. Join the waitlist and we'll show you how it works on your own consultations.