Multi-Modal Foundation · google
MedASR
A 105-million-parameter speech recognition model fine-tuned on clinical dictation for medical transcription.
What MedASR is
MedASR is an automated speech recognition model developed by Google to transcribe clinical audio into written text. General-purpose speech models often fail on complex clinical vocabularies, dosage notations, and specialist medical jargon. MedASR addresses this by training directly on medical audio to provide accurate transcription for healthcare documentation.
The model is designed as a foundational speech-to-text component for clinical software developers and researchers. It is tailored for dictation across clinical domains such as radiology, family medicine, and internal medicine, and can serve as an audio intake system paired with downstream generative models such as MedGemma.
Outputs from MedASR are intended as preliminary transcriptions requiring human verification and clinical correlation, rather than unverified inputs into immediate patient treatment or diagnostic decisions.
How it works
MedASR is built on the Conformer architecture and employs Connectionist Temporal Classification (CTC) for decoding speech into text. The model accepts single-channel 16 kHz audio waveforms and produces plain text transcriptions.
The model was initially pre-trained on speech using the full LibriHeavy dataset with JAX and ML Pathways running on Google TPU hardware (TPUv4p, TPUv5p, and TPUv5e). It was subsequently fine-tuned on more than 5,000 hours of de-identified, licensed physician dictations across multiple specialities. These training segments incorporated annotations for medical named entities, such as symptoms, medications, and clinical conditions. Transcriptions can be generated using greedy decoding or combined with an external language model via beam search.
Key facts
- Architecture
- Conformer with CTC (lasr_ctc)
- Parameter count
- 105M
- Input modality
- Mono-channel audio (16 kHz, int16 waveform)
- Output modality
- Text
- Release date
- 18 December 2025
- Pre-training data
- LibriHeavy training set
- Specialised training data
- >5,000 hours of de-identified physician dictation
What you can use it for
- Radiology Dictation Transcription
- Transcribing spoken radiologist narrative reports from diagnostic imaging reviews, achieving lower word error rates than general-purpose speech models.
- Outpatient and Inpatient Documentation
- Converting spoken notes from internal medicine and family medicine consultations into draft text for electronic health records.
- Speech Front-End for Clinical Language Models
- Providing an accurate audio-to-text pipeline that supplies transcribed clinical speech as text prompts to generative models such as MedGemma.
Run it here
We list this model but cannot run it yet: gated on the Hub — never run past a gate the user did not accept.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Outperforms larger general-purpose baselines such as Whisper v3 Large and Gemini 2.5 on medical dictation benchmarks.
- Maintains a compact 105-million-parameter footprint suitable for local execution and rapid inference.
- Trained with explicit supervision on clinical named entities including diagnoses, symptoms, and pharmaceuticals.
- Supports pairing with external n-gram language models to further reduce word error rates.
Watch out for
- Trained solely on English audio, with a training distribution skewed towards male speakers with American accents.
- Transcription accuracy declines when processing low-quality audio or recordings with ambient background noise.
- Exhibits variability when transcribing date formats, times, and durations due to anonymisation in the training data.
- May fail on non-standard drug names, emerging procedures, or terminology introduced within the past decade.
- Cannot be used for direct clinical management decisions without independent professional review and validation.
Inputs and outputs
You provide
Audio
WAV, MP3 or FLAC.
You get back
The model's completion.
Licence
Check the licence before commercial use · other
This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
Questions
Can I use MedASR commercially?
Check the licence before commercial use. This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
What audio input specifications does MedASR require?
The model expects single-channel (mono) audio sampled at 16 kHz in an int16 waveform format.
Can MedASR be fine-tuned on custom clinical datasets?
Yes, MedASR is designed to serve as an adaptable base model and can be fine-tuned using domain-specific audio and matched text transcripts.
Why does the model struggle with dates and durations?
The medical dictation corpus used for fine-tuning was rigorously de-identified to protect patient privacy, which removed or altered temporal metadata. This can be mitigated through external language model debiasing or targeted fine-tuning.
How does MedASR achieve its lowest word error rates?
The lowest word error rates are achieved by coupling MedASR with an external 6-gram language model using beam search with a beam size of 8.
How to cite it
Wu, Ke and Variani, Ehsan and Bagby, Tom and Riley, Michael. Last: Scalable Lattice-Based Speech Modelling in Jax. ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5, 2023. IEEE.
Written from MedASR's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.