fwasil162/whisper-medium-medical-pl-ggml

Whisper-medium medical-pl — ggml q5_0 (whisper.cpp)

Quantized ggml build of the ADMEDVOICE Polish medical fine-tune of openai/whisper-medium, packaged for whisper.cpp (v1.6.2 ggml format) on-device inference.

• File: ggml-medium-medical-pl-q5_0.bin (~514 MB, q5_0)
• SHA-256: a75091f8edb67e819e922d08d837b13537a849c2e66cb96efceb73bf6e6c953a
• Source checkpoint: msxksm/whisper-medium-medical-pl (Kaggle, Apache 2.0), LoRA variant natural_anonym_synth (the paper's best: WER 24.03 → 13.91 on the ADMEDVOICE test set), merged into the base model with PEFT merge_and_unload, converted with whisper.cpp convert-h5-to-ggml.py, quantized to q5_0.

Intended use

Research and engineering use. Transcription output is draft material requiring review by a qualified clinician — the model mishears drug names, dosages and negations, and a plausible-looking wrong number reads exactly like a right one.

This is not a medical device. Do not use it for diagnosis, for treatment decisions, or as a finished clinical record. Deploying it in patient care makes you responsible for qualifying the surrounding system under the applicable medical-device and AI rules.

Provenance & attribution

• Fine-tune and dataset by the ADMEDVOICE project — Gdańsk University of Technology (Multimedia Systems Dept.) and Medical University of Gdańsk; funded by Poland's National Centre for Research and Development (NCBR), project INFOSTRATEG4/0003/2022.
• Dataset: A Comprehensive Polish Medical Speech Dataset for Enhancing Automatic Medical Dictation, Scientific Data (2025) — corpus released under CC-BY.
• Base model: openai/whisper-medium (Apache-2.0).
• Fine-tuned checkpoint license: Apache 2.0 (as published on Kaggle).

This repackaging (merge + ggml conversion + quantization) is distributed under Apache-2.0, preserving the upstream licenses and attribution above.

Usage (whisper.cpp)

./main -m ggml-medium-medical-pl-q5_0.bin -l pl -f audio-16khz-mono.wav