Datasetpublic

MengjieChi/med-dictate

Med-Dictate — ASR Evaluation Dataset An evaluation dataset released by Corti ApS alongside the Symphony for Speech Recognition white-paper. Medical notes dictated by Corti team members and one contractor, with their written consent, in English, French, and German. Built for benchmarking automatic speech recognition (ASR) and related NLP systems on medical-domain audio. No real patient data. No PHI. No identifiable third-party content. Languages: en, fr, de How to… See the full description on the dataset page: https://huggingface.co/datasets/MengjieChi/med-dictate.

sourceHugging Faceotherupdated 2mo agoView on Hugging Face

1likes47downloads

Dataset Card

Med-Dictate — ASR Evaluation Dataset

An evaluation dataset released by Corti ApS alongside the *Symphony for Speech Recognition* white-paper. Medical notes dictated by Corti team members and one contractor, with their written consent, in English, French, and German. Built for benchmarking automatic speech recognition (ASR) and related NLP systems on medical-domain audio.

No real patient data. No PHI. No identifiable third-party content.

Languages: en, fr, de

How to use

python

from datasets import load_dataset

dataset = load_dataset("corti/med-dictate")           # English (default)
dataset = load_dataset("corti/med-dictate", "fr")     # French
dataset = load_dataset("corti/med-dictate", "de")     # German

Config: `default` / `en`

Medical notes in English.

Subset	Num. examples	Size	Duration
`test`	24	126.51 MB	1 h, 5 min, 49 s

Feature	Type
`id`	`Value(dtype='large_string')`
`transcription`	`Value(dtype='large_string')`
`transcription_formatted`	`Value(dtype='large_string')`
`medical_terms`	`Sequence(feature=Value(dtype='string'), length=-1)`
`medical_terms_formatted`	`Sequence(feature=Value(dtype='string'), length=-1)`
`formatting_terms`	`Sequence(feature=Value(dtype='string'), length=-1)`
`audio`	`Audio(sampling_rate=16000, mono=True, decode=True)`
`audio_path`	`Value(dtype='string')`

Config: `fr`

Medical notes in French.

Subset	Num. examples	Size	Duration
`test`	7	29.89 MB	15 min, 33 s

Feature	Type
`id`	`Value(dtype='large_string')`
`transcription`	`Value(dtype='large_string')`
`transcription_formatted`	`Value(dtype='large_string')`
`medical_terms`	`Sequence(feature=Value(dtype='string'), length=-1)`
`medical_terms_formatted`	`Sequence(feature=Value(dtype='string'), length=-1)`
`formatting_terms`	`Sequence(feature=Value(dtype='string'), length=-1)`
`audio`	`Audio(sampling_rate=16000, mono=True, decode=True)`
`audio_path`	`Value(dtype='string')`

Config: `de`

Medical notes in German.

Subset	Num. examples	Size	Duration
`test`	9	62.63 MB	32 min, 35 s

Feature	Type
`id`	`Value(dtype='large_string')`
`transcription`	`Value(dtype='large_string')`
`transcription_formatted`	`Value(dtype='large_string')`
`medical_terms`	`Sequence(feature=Value(dtype='string'), length=-1)`
`medical_terms_formatted`	`Sequence(feature=Value(dtype='string'), length=-1)`
`formatting_terms`	`Sequence(feature=Value(dtype='string'), length=-1)`
`audio`	`Audio(sampling_rate=16000, mono=True, decode=True)`
`audio_path`	`Value(dtype='string')`

Intended Use

1.Evaluating, benchmarking and testing ASR / NLP systems (including commercial benchmarking).
2.Academic and applied research.
3.Reproducing the results in the Corti white-paper.

Not Allowed

In plain English — please don't:

1.Use the data to train, fine-tune or otherwise develop any ML/AI model — generative or discriminative, voice or otherwise. This dataset is for evaluating pre-existing models only.
2.Clone, imitate or generate audio that sounds like any voice in the dataset.
3.Build a product that competes with commercial TTS or voice-cloning platforms.
4.Make deepfakes, impersonation content, or anything intended to deceive.
5.Use it for any clinical, diagnostic or treatment decision — this is not a medical device and has not been clinically validated.
6.Try to re-identify any contributor or any individual referenced in the source material.

These restrictions exist because the voice contributors consented only to evaluation-style use of their recordings.

Licence

Released under the Corti ASR Evaluation Dataset Licence, which combines:

1.CDLA-Permissive-2.0 — broad permission to use, modify and redistribute; and
2.a Corti Use Restrictions Addendum (RAIL-style) that imposes the restrictions listed above and flows them down to anyone you share the dataset with.

If the two parts ever conflict, the Addendum wins.

Full text: LICENCE.md · Governed by Danish law.

Voice Contributor Withdrawal

Voice contributors can withdraw their consent at any time. Upon Corti's notice, downstream users must stop distributing the affected recordings within 30 days. Withdrawal does not affect the lawfulness of prior use.

Citation

If you use this dataset, please cite:

Nix et al., "Symphony for Speech-to-Text: Supporting Real-Time Medical Voice Interfaces",
arXiv:2605.16545, 2026. https://doi.org/10.48550/arXiv.2605.16545

Contact

General correspondence: jdh@corti.ai Take-down requests or contributor withdrawals: privacy@corti.ai