Blog · Data for AI · Clinical speech
Can you train a commercial model on public clinical speech data?
Sometimes, but less often than the download page suggests. Public doctor-patient conversation datasets are few, mostly simulated, and frequently licensed for research or non-commercial use only. Before one goes into a training run for an ambient scribe or a medical speech recognizer, 3 things need checking: what the data actually is, what its license allows, and whether the permissions behind it cover AI training.
Check 1
How much public clinical conversation data is there?
Very little, and most of it is not recorded in a clinic. A 2026 survey of datasets for ambient clinical scribes counted the patient-clinician conversation datasets it could find and classified them by origin.
13
patient-clinician conversation datasets identified
7
simulated consultations
4
real, but text from online medical forums
1
from real clinical encounters, and not public
None of this means simulated data is a poor choice. Well-designed simulated consultations, recorded by practicing clinicians from case briefs, are a legitimate and often the most practical route. But "public clinical conversation dataset" usually means simulated audio or forum text, and a buyer should know which before comparing it with anything else. The differences between the 3 classes, and when each one fits, are covered in our guide Real, simulated, or synthetic: how to source clinical conversation data for AI.
Check 2
Does the license allow commercial training?
Often it does not. MeDial-Speech, released in May 2026 with more than 111 hours of doctor-patient and robot-patient consultations, is free of charge for non-commercial purposes. That makes it useful for research and of no use as training data for a commercial scribe, however good the audio is.
The problem runs deeper than the license on the download page. An audit of nearly 4,000 public text, speech, and video datasets (ICLR 2025) found that fewer than 33% of the datasets were restrictively licensed, while more than 80% of the source content inside them carried non-commercial restrictions. A permissive license on a dataset does not guarantee that the material it was built from allows commercial use.
THE GAP
Dataset license vs. source content
- Under 33% of audited datasets were restrictively licensed.
- Over 80% of their source content carried non-commercial restrictions.
Longpre et al., Bridging the Data Provenance Gap Across Text, Speech, and Video, ICLR 2025.
Check 3
Do the permissions behind the data cover AI training?
A license is only as good as the rights of whoever grants it. For clinical speech, that chain has 4 links: the basis on which the recording was collected, how identifiable information was removed from the transcript and from the voice, whether the seller may license the data onward, and whether the original consent covers commercial AI training rather than only the purpose the recording was made for. In the EU, health data is a special category under Article 9 of the GDPR, and a voice can identify a speaker even after names are removed.
Most failures sit in the last 2 links, and nothing in the audio reveals them. If a vendor answers a rights question with a sample, the question has not been answered. Our reference page on data provenance and licensing sets out the permission chain, the 8 license terms that decide what a dataset can be used for, and 8 questions to send any data vendor.
What to do instead
Where should a commercial clinical speech program get its data?
Use public datasets for what their licenses allow, such as research and benchmarking, and source training data whose rights you can document. Start with the page that matches where you are.
Data provenance and licensing
How to check where any training dataset came from and what you may do with it.
Read the reference GuideReal, simulated, or synthetic
Which class of clinical conversation data fits training, evaluation, and augmentation, and how to test it.
Read the guide DatasetsClinical speech data
Doctor-patient conversation datasets with a provenance label on every hour and a documented permission chain.
See the datasetsSOURCES
Evaluating Ambient Clinical Scribes in India: The Need for Multilingual Real-World Clinical Conversation Data (2026), Table 1. arxiv.org/abs/2609.17355
Cuayáhuitl, H. and Jang, G. (2026). A dataset of robot-patient and doctor-patient medical dialogues for spoken language processing tasks (MeDial-Speech). arxiv.org/abs/2605.26747
Longpre, S. et al. (2025). Bridging the Data Provenance Gap Across Text, Speech, and Video. ICLR 2025. arxiv.org/abs/2412.17847
Regulation (EU) 2016/679 (GDPR), Article 9. eur-lex.europa.eu/eli/reg/2016/679/oj
Need clinical speech data you can train on?
Tell us the languages, settings, and volumes you need. We reply in writing with the provenance class, the license scope, and sample recordings.