RadiOtter/RadiOtter-ASR-v1

RadiOtter ASR v1

RadiOtter ASR v1 is an English speech-to-text model designed for radiology dictation. It is a fine-tuned version of Whisper large-v3-turbo, provided in CTranslate2 format for use with Faster Whisper.

Synthetic data was used in the model's preparation.

Intended coverage

Modalities: CT, MR, XR, US and MG.

Subspecialties: Abdomen, Neuro, Breast, MSK, Vascular, Nuclear Medicine (NM), OB-GYN, Spine and Chest.

Performance comparison

Word error rate (WER) measures transcription differences; lower is better.

RadiOtter’s figure is reported from a separate evaluation. The remaining figures are from our comparison run; the results are not directly comparable without matching evaluation details and are not clinical validation.

Model	Normalized WER ↓
RadiOtter ASR v1	2.47%
Qwen3-ASR 1.7B	6.66%
Google MedASR	7.95%
Cohere Transcribe	10.90%
IBM Granite Speech 5.0 Turbo CTC	13.61%
NVIDIA Nemotron 3.5 ASR Streaming	22.77%

Usage

For a ready-to-run local dictation app with live transcription and audio-file support, see the RadiOtter ASR GitHub repository. Its README includes setup and usage instructions.

Install Faster Whisper:

pip install faster-whisper

Load the CTranslate2 model from its repository:

from faster_whisper import WhisperModel

model = WhisperModel(
    "RadiOtter/RadiOtter-ASR-v1",
    device="cpu",
    compute_type="int8",
)

segments, info = model.transcribe("audio.wav", language="en", beam_size=5)
print(" ".join(segment.text.strip() for segment in segments))

Intended use and limitations

RadiOtter is intended for developers and researchers exploring radiology speech recognition and dictation workflows. Transcriptions may contain errors, including in negation, laterality, anatomy, numbers and measurements. A qualified professional should review all output before it is used in a clinical workflow. Performance across modalities, subspecialties, accents, recording conditions and real-world dictation requires further validation.