Resources
Open test audio for Spanish clinical speech is close to nonexistent. We searched Hugging Face, Zenodo and the research literature, and found exactly one openly licensed set of Spanish doctor-patient recordings with transcripts that you can download today.
We converted it into the same format as our English audio set: 90 Spanish visits, about 4 hours, each with a reference transcript. It's free on Hugging Face. Sully maintains it and sells a clinical speech and notes API, so the code below uses Sully, but the audio works with any vendor.
One thing to know up front. These voices are AI-generated, not people. That changes what a score on this set means, and the sections below say how.
Key Takeaways
90 recordings, about 4 hours, CC BY 4.0. Commercial use is allowed with credit to the original authors.
The voices are synthetic. The authors generated the audio with Google Gemini text to speech, then degraded it to 8 kHz to sound like a clinic. 23 of the 90 scripts are based on real anonymized consultations.
It's strong on vocabulary. Spain Spanish across 12 specialties, with drug names, doses and Spain-specific brands like Adiro and Nolotil.
It's weak on realism. Clean speech, few filler words and no accent variety. Use it to test medical terms and catch regressions, not to predict accuracy in a real clinic.
What's in the Audio Set
The recordings come from one published research dataset. We converted it into the format of our English set: 16 kHz mono MP3, with transcripts that use the same Clinician:, Patient: and Family member: labels, so the same scoring code runs on both.
Source | Recordings | Audio | Setting | Extras | License |
|---|---|---|---|---|---|
TEME-v1 (Villarroya Martínez et al.) | 90 | About 4 hours (1.4 to 5.1 minutes each) | Spain Spanish. 29 cardiology, 16 primary care, 8 neurology, 8 pulmonology, plus emergency, orthopedics, geriatrics, pediatrics, endocrinology, oncology, gynecology and rheumatology | Specialty, case title and summary | CC BY 4.0 |
The 90 visits add up to about 37,000 words, with a median of about 370 words per visit. That's much shorter than the English set, where a median visit runs to about 1,300 words.
The dataset comes from the TEME project, an evaluation framework for Spanish medical speech recognition presented at AAMAS 2026. The authors built it to show that word error rate misses the errors that matter clinically, which is the same problem this article is about.
What it sounds like
Here's a moment from one of the cardiology recordings (teme-005), as written in the reference transcript:
Clinician: Eso refuerza la sospecha. Si el dolor empeora aquí, avísenos inmediatamente. También podríamos administrarle nitroglicerina sublingual, 0,4 mg, si reaparece el dolor opresivo.
Look at "0,4 mg". Spanish writes decimals with a comma, so a speech API that returns "0.4 mg" or "cero coma cuatro miligramos" is right and will still be scored wrong. Across the set there are more than 100 doses written as numbers with "mg", so number handling decides a real share of your score.
How to Download the Recordings
The whole set is about 82 MB, so you can pull all of it. The metadata file holds every transcript:
Each line of audio/metadata.jsonl has the file_name, encounter_id, specialty, duration_seconds, license and the full transcript. If you also want the case summary and turn-by-turn speaker roles, load the encounters config:
How to Measure Word Error Rate on Spanish Medical Audio
Word error rate (WER) counts the words a system got wrong, added or dropped, divided by the words in the reference transcript. In Spanish it's even easier to get a misleading number, because accents, inverted question marks and decimal commas all count as differences if you don't normalize them.
Send the audio to a speech API
Sully's transcription endpoint is asynchronous: you upload a file, get an ID back, and poll until it's done. These recordings are Spain Spanish, so pass es. For Latin American audio the endpoint also accepts es-419.
Normalize both sides, then score
Strip the speaker labels, punctuation (including ¿ and ¡) and filler words from both sides before you compare. Then decide on accents. Keep them if you care whether "está" and "esta" come back right, or fold them if you only care about the words.
Numbers written as words are the next problem. One system writes "cinco días", another writes "5 días", and both are right. Pick one form and convert both sides to it before you trust a comparison between vendors.
Look past the average
An overall WER hides the errors that matter. Missing "pues" costs the same as turning "enalapril" into "en a la pril", and only one of those changes a chart.
This set is useful here because the drug names are dense. Paracetamol, ibuprofeno, enalapril, atorvastatina, salbutamol, omeprazol and metformina each appear 12 times or more, and many visits pair a drug with a dose and a frequency. Pull the drug and dose phrases out of each reference transcript and score them on their own. The article on WER and drug names covers how to build that term-level check.
Testing Speaker Labels
Speech APIs return numbered speakers, like speaker: 0, not roles, so you have to map speakers to clinician and patient yourself. This set doesn't include turn timestamps, so it can't tell you which words landed on the wrong role.
What it can do is check your mapping at the turn level. Every visit has an ordered list of turns with roles, so you can test whether your mapping step names the clinician correctly. Four pediatric visits have a parent answering for the child, labeled Family member, which is a good test for any two-speaker assumption.
For timed turns, use the PriMock57 part of our English audio set. The speaker diarization guide covers mapping labels to roles.
What This Audio Won't Tell You
Know where this set differs from your users before you trust a number from it.
The voices aren't human. Text to speech doesn't stumble, mumble or talk over the other person. We counted only a handful of hesitation sounds in about 37,000 words. Expect real visits to score worse.
It's one variety of Spanish. The scripts use Spain Spanish ("vale" appears 115 times, along with "médico de cabecera" and brands like Adiro). There are no Mexican, Caribbean or other Latin American speakers, and no patients switching between Spanish and English. The multilingual transcription guide covers that case.
The source audio is 8 kHz. The authors degraded it on purpose to sound like a clinic or a phone line. We re-encoded it to 16 kHz to match the English set, which doesn't restore the lost detail.
The scripts have a few slips. One endocrinology visit (teme-066) discusses polycystic ovary syndrome with a male-voiced patient, a change the authors marked in its title. In teme-009, a female patient describes herself as "desesperado". Keep that in mind if you read transcripts closely.
It leans cardiology. 29 of the 90 visits are cardiology, so cardiac drugs are well covered and specialties like dermatology are missing.
Use this set to shortlist vendors on Spanish medical vocabulary and to catch regressions when you change models. Then confirm on recordings from your own setting, under a BAA, before you go live. The HIPAA medical transcription API comparison lists which vendors sign one.
How Sully Fits
The Sully API covers the steps after the audio. Recorded transcription takes Spanish as es or es-419 and returns word timings and optional speaker labels. For live audio, streaming transcription returns partial and final transcripts during the visit, and lets you boost the words that matter, such as drug names.
The transcript then goes to Note Generation with your template, and Text to JSON turns the note into structured fields for the EHR.
To try it, create an account and get keys on the Sully API page, then run the scripts above on ten recordings. It takes a few minutes, and you'll have your first WER baseline on Spanish medical audio.
Frequently Asked Questions
Is there a free dataset of Spanish doctor-patient conversation audio?
Yes. The Spanish Open Clinical Conversations set on Hugging Face has 90 Spanish doctor-patient recordings, about 4 hours, with reference transcripts. It's a converted copy of the TEME-v1 dataset (CC BY 4.0).
Are the recordings real patients?
No. The authors wrote the dialogues, 23 of them based on real anonymized consultations, had a physician validate them, and generated the voices with Google Gemini text to speech. There's no patient audio in the set.
Can I use this audio in a commercial product?
Yes. CC BY 4.0 allows commercial use as long as you credit the original authors. The attribution file has the wording, and each row carries its license.
Is it Spain Spanish or Latin American Spanish?
Spain Spanish. The vocabulary, drug brands and phrasing are from Spain. If your users are in Mexico, the US or elsewhere in Latin America, treat scores from this set as a vocabulary check only.
Can I train a speech model on it?
The license allows it, but 4 hours of synthetic speech is far too little for training and would teach a model the quirks of one text to speech system. It's better used as a held-out test set.
Is there an English version?
Yes. The English audio set has 329 human role-played recordings, about 60 hours, in the same format. The same scoring code runs on both.
Sources
Resources

