medasr-onnx

MedASR ONNX — BizenLabs

ONNX-exported variants of google/medasr for local, offline medical speech recognition.

Published at: bizenlabs/medasr-onnx

Model Variants

Quick Start

import numpy as np
import onnxruntime as ort
from transformers import AutoProcessor

processor = AutoProcessor.from_pretrained("google/medasr", trust_remote_code=True)
session = ort.InferenceSession("model-fp32.onnx", providers=["CPUExecutionProvider"])

# audio: float32 numpy array, mono, 16 kHz
encoded = processor(audio, sampling_rate=16000, return_tensors="np", padding=True)
feed = {"input_features": encoded["input_features"]}
if "attention_mask" in encoded:
    feed["attention_mask"] = encoded["attention_mask"]

logits = session.run(None, feed)[0]
token_ids = np.argmax(logits, axis=-1)
text = processor.batch_decode(token_ids)[0]

Source Model

Conversion Details

• Exported with torch.onnx.export using legacy TorchScript path (dynamo=False)
• ONNX opset version: 17
• Dynamic axes: batch + sequence (supports variable-length audio)
• FP16: mixed-precision with LayerNormalization, Softmax, ReduceMean, Attention ops kept in FP32
• INT8: dynamic quantization on MatMul/Gemm ops
• All variants validated against FP32 baseline (logit-level max_abs_diff + text-level WER)

See CONVERSION.md for the full process, key decisions, and reconversion checklist.

Directory Contents

Setup

cd tools/medasr-onnx
python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.lock.txt

Login to Hugging Face (required for gated model access):

huggingface-cli login

Convert

python convert_medasr.py \
  --model-id google/medasr \
  --version 2026.02.20 \
  --output-dir ./artifacts

The script will:

1. Export FP32 baseline and validate (structure, dynamic axes, runtime load)
2. Build FP16 variant — auto-rejected if logit divergence or WER exceeds threshold
3. Build INT8 variant with accuracy gate
4. Write manifest.json with SHA256 checksums for all passing variants

To skip FP16 entirely:

python convert_medasr.py \
  --model-id google/medasr \
  --version 2026.02.20 \
  --output-dir ./artifacts \
  --skip-fp16

Upload to HuggingFace

Preview what will be uploaded:

python upload.py --version 2026.02.20 --dry-run

Upload all passing variants:

python upload.py --version 2026.02.20

Upload specific variants only:

python upload.py --version 2026.02.20 --variants fp32 int8

Reconverting (New Model Release)

When Google releases a new MedASR version:

1. Update the transformers pin in requirements.lock.txt if the model card specifies a new commit
2. Run conversion: python convert_medasr.py --model-id google/medasr --version YYYY.MM.DD
3. Review the console output for any rejected variants
4. Upload: python upload.py --version YYYY.MM.DD

See CONVERSION.md for the full reconversion checklist.

License

ONNX conversion code: Apache 2.0 Model weights: subject to google/medasr license (Apache 2.0 code + HAI-DEF weights)