B-CON AI · hearing · Speech to text
B-CON Ears
Hears Nepali the way it is spoken in a clinic.
A speech recognition model trained for Nepali, including the English drug names, test names and numbers that fill a real consultation. It turns a doctor and patient talking into text in about a second, on our own servers, so no audio ever leaves the hospital.
Capabilities
What it does.
- 01Transcribes Nepali conversation in Devanagari, with English clinical terms kept intact
- 02Separates what the doctor said from what the patient said
- 03Understands dose patterns such as 1-0-1 and dates in Bikram Sambat
- 04Runs on CPU for small hospitals and GPU for large ones
- 05Works through a normal browser microphone, nothing to install
Specification
How it is built and served.
- Language
- Nepali (Devanagari), code-switched English medical vocabulary
- Architecture
- Conformer encoder, hybrid transducer and CTC decoding
- Latency
- About one second per sentence
- Serving
- Self-hosted, CPU or GPU, streaming over WebSocket
- Data residency
- Audio processed and discarded inside the hospital network
- Trained on
- Nepali speech corpora plus a clinical vocabulary layer, no patient data
Why we trained it
Every other speech model treats Nepali as an afterthought. Ours treats it as the point.
Doctors in Nepal switch between Nepali and English mid-sentence: the complaint in Nepali, the drug in English, the dose as 1-0-1. General models drop half of it. B-CON Ears was built around that exact pattern.
What it does inside the EHR
- 01
Consultation transcription
The whole visit is captured as text while the doctor keeps looking at the patient, not the screen.
- 02
Dictation anywhere
Ward rounds, radiology reports, discharge summaries and nursing notes all take voice.
- 03
Voice search
Find a patient, a drug or a test by saying it, in Nepali.
- 04
Patient app
Patients talk to the B-CON PHR assistant in Nepali and get answers from their own record.
The other senses