fonendo-bench is released under the Omniloy fonendo-bench evaluation license: you may use it only to evaluate and research speech-to-text systems, you may not redistribute the audio, and you may not use the audio to train, fine-tune or adapt any model. Access is granted by hand. Fill in the form below and also email info@omniloy.com with your name, your organisation and your intended use; we answer every request.

fonendo-bench: clinical subsets

Spanish clinical dictation for evaluating speech-to-text systems. This repository holds the two clinical subsets of fonendo-bench, a reproducible Spanish clinical speech-to-text benchmark published by Omniloy. The benchmark code, the tools that rebuild the public (real speech) subsets and the leaderboard live in the GitHub repository; this dataset contains only the clinical audio, its reference transcripts and its metadata.

config	clips	sentences	hours	purpose
`clinical_test`	300	242	0.64	the reported clinical test set
`clinical_dev`	60	60	0.13	smoke tests and runner development; never reported

Each config has a single split, test. The sentences of clinical_dev are disjoint from those of clinical_test.

Access. The dataset is gated. Request access on this page and email info@omniloy.com with your name, your organisation and your intended use. Use is limited to evaluation and research (see License).

What is in it

Sentences. Fictitious Spanish clinical sentences written for the benchmark in the register of clinical notes, discharge summaries and test orders: drug names with doses, diagnoses, imaging and laboratory tests, abbreviations (TAC, RMN, PCR, ...), numbers and units. They describe no real patient and contain no personal data.

Voices. Every sentence was synthesised with text-to-speech: ElevenLabs (a commercial service; 9 Spanish voices, from Spain and Latin America) and Kokoro (Kokoro-82M, an open-weights model under Apache-2.0; 2 voices from Spain). Synthetic speech is more regular than a clinician dictating in a hospital, so absolute error rates here are likely optimistic compared with real dictation; the benchmark's real-speech subsets (in the GitHub repository) are the counterpart.

Acoustic conditions. About half of the clips are clean (the TTS output, unperturbed). The other half are degraded with a random combination of controlled perturbations, recorded per clip:

additive noise at a target signal-to-noise ratio of 5, 10, 15 or 20 dB: babble (6 overlapping speakers from MUSAN speech), environmental noise (MUSAN noise), or synthetic white, pink or brown noise;
room reverberation with simulated room impulse responses (OpenSLR SLR28) on some clips;
speed perturbation (0.92 to 1.08) and pitch shift (plus or minus 0.5 or 1.2 semitones) on some clips;
a round trip through a narrow-band speech codec on some clips: AMR-NB (5.9 or 10.2 kbit/s) or Opus (8 or 16 kbit/s).

Sampling. clinical_test is balanced 50/50 between two families of medical terms (drug_or_diagnosis and diagnostic_test), spread across 19 specialties, and balanced across the four cells TTS engine x condition (72 to 77 clips each). A sentence appears at most once in the drug/diagnosis family and at most twice (in two different cells) in the diagnostic-test family; text_id groups the renditions of a sentence. clinical_dev has one clip per sentence and covers internal medicine, emergency medicine and traumatology (59 drug/diagnosis clips, 1 diagnostic-test clip).

Summary of `clinical_test`


clips / distinct sentences	300 / 242
total duration	38.2 min (clips of 3.7 to 13.6 s, median 7.5 s)
reference words	5,615
gold term occurrences / distinct terms	385 / 272
TTS engine	ElevenLabs 149, Kokoro 151
voice region	Spain 255, Latin America 45
voice gender	female 160, male 140
condition	clean 147, degraded 153
specialties	traumatology 61, emergency medicine 61, internal medicine 60, neurology 12, gastroenterology 12, pulmonology 12, cardiology 11, endocrinology 11, nephrology 10, oncology 8, radiology 8, rheumatology 8, urology 6, primary care 6, hematology 4, nuclear medicine 4, pediatrics 2, vascular surgery 2, general surgery 2

Fields

column	type	description
`clip_id`	string	stable clip identifier, the key used by fonendo-bench results files
`audio`	audio	16 kHz, mono, 16-bit, FLAC (lossless)
`text`	string	reference transcript as written (casing, punctuation, digits and units kept); scoring normalizes it
`terms`	list of string	the gold medical terms spoken in the clip (see below)
`text_id`	string	sentence identifier; several clips can share a sentence
`term_family`	string	`drug_or_diagnosis` or `diagnostic_test`
`specialty`	string	medical specialty of the sentence
`tts_engine`	string	`elevenlabs` or `kokoro`
`voice_region`	string	`spain` or `latin_america`
`voice_gender`	string	`female` or `male`
`condition`	string	`clean` or `degraded`
`snr_db`	float	target signal-to-noise ratio of the added noise; null when clean
`noise_type`	string	`babble`, `environmental`, `white`, `pink` or `brown`; null when clean
`reverb`	bool	room reverberation applied
`codec`	string	`amr_nb` or `opus` when the clip went through a codec, else null
`codec_bitrate`	int	codec bit rate in bit/s, else null
`speed`	float	speed factor (1.0 = unchanged)
`pitch_semitones`	float	pitch shift in semitones (0.0 = unchanged)
`duration_s`	float	duration in seconds

The `terms` field

terms lists the medical terms that are actually spoken in the clip (drug names, diagnoses, imaging and laboratory tests), written as they appear in text. Most clips have one term, some two or three. They are the ground truth for the benchmark's term metrics:

term recall: the share of gold-term occurrences in the normalized reference whose every word is aligned as correct in the system's transcript (word-level alignment of reference and transcript; all or nothing per term). The denominator counts the occurrences found verbatim in the normalized reference: 384 of the 385 in clinical_test;
term word error rate (B-WER) and other word error rate (U-WER): the word error rate restricted to the words of the gold terms (function words such as "de" excluded) and to all other words.

Terms are used only for scoring. Every system on the fonendo-bench leaderboard except the results-only rows transcribes the audio in its default configuration (no custom vocabulary, keyterms or context prompt; Spanish selected where the system allows it; instruction-following models get only the fixed transcription instruction they need) and never sees terms, text or any other field. Results-only rows of the leaderboard are evaluated by their owner and state their conditions: the published OmniScribe 2 row used context from the patient's record that included the medical terms spoken in each clip, so its clinical numbers are a best case: a real record may not list every term a clinician says.

Usage

With fonendo-bench

pip install "fonendo[whisper] @ git+https://github.com/Omniloy/fonendo-bench"
huggingface-cli login            # or: export HF_TOKEN=...  (an account with access)
fonendo run --model whisper_large_v3 --subset clinical_test
fonendo score --subset clinical_test --hyps results/raw/whisper_large_v3/clinical_test.jsonl

fonendo.data.load_subset("clinical_test") returns the clips as 16 kHz float32 arrays with text, terms and the metadata; see the repository for the runners, the scoring rules and the leaderboard.

With `datasets`

from datasets import Audio, load_dataset
import io, soundfile as sf

ds = load_dataset("Omniloy/fonendo-bench", "clinical_test", split="test")
ds = ds.cast_column("audio", Audio(decode=False))   # decode with soundfile, no torchcodec needed
row = ds[0]
audio, sr = sf.read(io.BytesIO(row["audio"]["bytes"]), dtype="float32")
print(row["clip_id"], sr, row["text"], row["terms"])

With decode=True (the default) recent datasets releases decode the audio with torchcodec.

Intended use and limitations

Use: comparing speech-to-text systems on Spanish clinical vocabulary, with the protocol of the fonendo-bench repository. Report clinical_test only; use clinical_dev for smoke tests and debugging, never to tune a system you then report.
Synthetic speech: TTS voices do not reproduce the disfluencies, hesitations, speaking rate or microphone conditions of real clinicians. Treat results as a controlled test of medical vocabulary, not as an estimate of accuracy in a hospital.
Size: 300 clips and 242 sentences. Differences between systems of a point or two of word error rate are often within the confidence interval; fonendo-bench reports 95% bootstrap intervals and paired comparisons for that reason (use sentence-level resampling via text_id for the most conservative intervals).
Coverage: 19 specialties, but three of them (traumatology, emergency medicine, internal medicine) provide 60% of the clips, and most voices are from Spain.
Not for training. The license forbids training or adapting models on the audio; a test set seen in training no longer measures anything.
Not a medical device. Nothing here validates a system for clinical use.

License

The dataset is distributed under the Omniloy fonendo-bench evaluation license (LICENSE.md): evaluation and research use only, no redistribution of the audio, and no training, fine-tuning or adaptation of models on the audio. Publishing scores, aggregate analyses and short transcript excerpts for illustration is allowed and encouraged, with a citation. The benchmark code in the GitHub repository is Apache-2.0.

The degraded clips use noise from MUSAN (Snyder, Chen and Povey, 2015; CC BY 4.0) and room impulse responses from OpenSLR SLR28 (Ko et al., 2017; Apache-2.0).

Public demo clips. 13 clinical_test clips and their reference transcripts are public: Omniloy plays them, as MP3, on the fonendo-bench showcase page (docs/audio/ in the GitHub repository, © Omniloy, for listening only). Every other clip is available only through this gated dataset.

Citation

@misc{omniloy2026fonendobench,
  title        = {fonendo-bench: a reproducible {S}panish clinical speech-to-text benchmark},
  author       = {{Omniloy}},
  year         = {2026},
  howpublished = {\url{https://github.com/Omniloy/fonendo-bench}},
  note         = {Clinical subsets: \url{https://huggingface.co/datasets/Omniloy/fonendo-bench}}
}

Contact

info@omniloy.com

Downloads last month: 28

Total file size:

50.4 MB

Request access to fonendo-bench