Speech Recognition (ASR)

VoxMedic

The most accurate Polish medical speech recognition model. Up to 55% fewer transcription errors than Whisper — dictate medical documentation quickly, accurately, and securely.

100% On-PremiseLocal processingWhisper large-v3-turbo
Request a VoxMedic demo
RECORDING  ·  medical documentation dictation (Polish audio)
00:18.4  |  16 kHz  |  PL
VoxMedicASR
Pacjentka lat 67

Model Description

VoxMedic is the most accurate automatic speech recognition (ASR) model specialized in Polish medical speech. Built on the Whisper large-v3-turbo architecture and fine-tuned on thousands of medical recordings, VoxMedic handles clinical terminology, drug names, and specialist vocabulary — the biggest pain points of general-purpose speech-to-text systems.

The results speak for themselves: on medical speech recordings VoxMedic reduces transcription errors (WER) by up to 55% compared to the base Whisper model — from 17.07% down to just 7.64%. Crucially, thanks to a proprietary anti-forgetting training recipe (general-domain data replay, knowledge distillation, and partial LoRA weight merging), the model retains full accuracy on general Polish speech and improves on the base model by up to 42% on out-of-domain recordings.

VoxMedic was built for real clinical work: dictating discharge summaries and test descriptions, transcribing visits, and automating medical documentation. With only 0.8B parameters, the model runs fast even on a single GPU and can be deployed entirely within your facility's infrastructure.

Patient recordings never leave your facility

VoxMedic runs fully locally (on-premise) on your organization's infrastructure. Physician voice data and patient information are never sent to an external cloud, ensuring full GDPR compliance and medical data security.

VoxMedic vs Whisper — transcription comparison

The same report produced by two models. Whisper fails on drug names — VoxMedic transcribes them flawlessly.

VoxMedicWER 7.64%
Pacjent zgłasza
Whisper large-v3-turboWER 17.07%
Pacjent zgłasza
correct medical terminologytranscription error

Content WER on Polish medical speech recordings (admed dataset, human-read). VoxMedic: 7.64% vs base Whisper large-v3-turbo: 17.07% — 55% fewer errors.

Key Features

Best accuracy on medical speech
Up to 55% fewer transcription errors (WER 7.64% vs 17.07%) than base Whisper large-v3-turbo on Polish medical recordings
Understands clinical terminology
Accurately transcribes drug names, diagnoses, abbreviations, and specialist medical vocabulary
No loss on general speech
Proprietary anti-forgetting recipe (data replay, knowledge distillation, partial merge) preserves general-domain accuracy
Fast turbo architecture
Just 0.8B parameters — real-time transcription even on a single GPU
Fully local operation
On-premise deployment within your facility's infrastructure — recordings never reach the cloud
Easy integration
REST API plus full compatibility with the Whisper ecosystem and the transformers library

Supported Formats

Dictation
Real-time dictation of discharge summaries, test descriptions, and clinical notes
Audio files
Transcription of WAV, MP3, FLAC, and other popular audio formats
Mobile
Recordings from dictaphones and mobile apps, even in difficult acoustic conditions
HIS/EHR integration
Streaming transcriptions directly into hospital systems via API

Use Cases

  • Dictating medical documentation — discharge summaries, interviews, test descriptions
  • Transcribing medical visits and telemedicine consultations
  • Voice-based reporting of imaging studies (radiology, ultrasound)
  • Creating clinical notes by voice during patient examination
  • Digitizing archival medical recordings
  • Feeding NLP systems (e.g. MedalioNER) with structured transcriptions

Business Benefits

Less time at the keyboard
Physicians spend up to a third of their working time on documentation. Dictation with VoxMedic cuts that time dramatically.
Fewer documentation errors
Medical specialization means correct spelling of drugs, doses, and diagnoses — no manual corrections.
Better staff experience
Less burnout and more time for patients thanks to automating tedious documentation.
Predictable costs
A local model with no per-minute fees — a one-time deployment instead of growing cloud bills.

FAQ – Frequently Asked Questions

API Integration

Models are available as a REST API service or for local deployment within hospital infrastructure. We provide documentation and integration support.

voxmedic_example.sh
bash
curl -X POST https://api.medalion.tech/v1/transcribe \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "file=@visit_recording.wav" \
  -F "model=VoxMedic" \
  -F "language=pl"

# Response:
{
  "text": "Pacjentka lat 67 przyjęta na oddział
    kardiologii z powodu nasilającej się
    duszności. Włączono furosemid 40 mg
    oraz ramipril 5 mg.",
  "duration": 18.4,
  "language": "pl"
}
Get started today

Test VoxMedic on recordings from your facility

Contact us to get API access or discuss an on-premise deployment.