This dataset is derived from the ESLO and LangAge French speech corpora and contains recordings of (anonymised) speakers. Access is granted manually by the maintaining organisation after review of your request. By requesting access you agree to: (1) use the data for non-commercial academic research only, in accordance with the CC BY-NC-SA 4.0 licence and the terms of the underlying ESLO and LangAge corpora; (2) not attempt to re-identify speakers or use the data to the detriment of any individual; (3) not redistribute the audio or transcripts outside your research team; and (4) cite this dataset and the source corpora in any resulting work.

FrWhisper Training Dataset (ESLO + LangAge)

French conversational-speech dataset used to fine-tune FrWhisper, a Whisper Large-V3 model adapted to spoken French, including interjections, hesitations, and other natural speech phenomena. It combines material from two French corpora, ESLO and LangAge, segmented into utterance-level clips paired with their transcriptions.

This release provides transcripts + audio only (16 kHz mono); it does not include precomputed log-mel features.

Dataset summary

Split	Segments	Audio (h)
train	77,081	~42.4
test	19,262	~10.6
total	96,343	~53

Language: French (fr)
Sampling rate: 16 kHz, mono, WAV (16-bit PCM)
Source recordings: 111 (55 ESLO + 56 LangAge interviews)
Domain: sociolinguistic interviews and conversational spoken French

Note: the split is at the segment level (random 80/20, seed 42), so the same recordings/speakers can appear in both train and test. It is not a speaker-disjoint split, so keep this in mind for evaluation.

Data fields

Column	Type	Description
`audio`	`Audio` (16 kHz)	The audio segment (decodes to `{array, sampling_rate}`).
`sentence`	`string`	French transcription of the segment.
`corpus`	`string`	Source corpus: `ESLO` or `LangAge`.
`source_file`	`string`	Source recording id (e.g. `ESLO2_ENT_26-46_1008`, `a005a`).
`speaker`	`string`	Within-recording tier label (e.g. `spk2`). This is not a globally unique speaker id; the interviewer tier was removed during preprocessing.
`segment`	`string`	Interval index of the segment within the recording.

Example clips (disfluent speech)

The dataset preserves natural speech phenomena. The examples below (LangAge and ESLO, different speakers) illustrate hesitations (euh), repetitions, spoken numbers, and interjections (ah, oh, hein). Audio players render for users logged in to the Hugging Face Hub with access to this repository.

Corpus	Transcript	Shows
LangAge	moi j' ai commencé à travailler euh euh c' était la guerre en mille-neuf-cent-quarante-deux hein	hesitation, number, interjection
LangAge	ah oui je lis euh deux ou trois journaux par par jour	interjection, hesitation, number, repetition
ESLO	de de de tout petit euh bah à partir de cinq ans hein	repetition, hesitation, interjection, number
ESLO	euh moins depuis avant quatorze oui c' était très très marqué	hesitation, number, repetition

Source corpora

ESLO: Enquêtes SocioLinguistiques à Orléans (ESLO1 and ESLO2), Laboratoire Ligérien de Linguistique (LLL), Université d'Orléans / CNRS. http://eslo.huma-num.fr/index.php Copyright (c) 2012 Université d'Orléans / LLL. Freely available for non-commercial use under a Creative Commons license.
LangAge: corpus of conversational French focused on older speakers, Universität Potsdam. https://www.uni-potsdam.de/en/langage-corpora/index

Please respect and cite the terms of the original corpora in addition to this derived dataset.

Preprocessing

The same core pipeline was applied to both corpora:

Audio resampled to 16 kHz mono (ESLO originals are 44.1/48 kHz stereo).
TextGrid intervals sliced into utterance-level segments paired with their transcriptions.
Cleaning: removal of empty/silent and very short (<100 ms) segments, bracketed non-speech markers, and the interviewer tier (spk1); a few problematic recordings were excluded.

In addition, ESLO transcriptions were harmonised to the LangAge annotation conventions: laughter and noise markers were normalised (for example (rire) to (rir), (sonnerie) to (bru)), idiosyncratic or unintelligible parenthesised tokens were removed, speaker-overlap tags were stripped, and apostrophe spacing was normalised (with a set of protected multi-word forms such as aujourd'hui and d'accord).

Usage

This is a gated dataset: you must be logged in to the Hugging Face Hub and have been granted access.

from datasets import load_dataset

ds = load_dataset("aihpi/FrWhisper-dataset")
print(ds)
sample = ds["train"][0]
sample["audio"]      # {'array': ndarray, 'sampling_rate': 16000}
sample["sentence"]   # French transcription

Intended use and limitations

Intended use: non-commercial academic research on French ASR and spoken French (fine-tuning and evaluation of speech models, linguistic analysis).
Out of scope: commercial use (prohibited by the NC licence term); speaker identification or any use intended to harm individuals.
Limitations: spontaneous-speech transcription conventions (hesitations, truncations); segment-level (non-speaker-disjoint) train/test split.

Personal and sensitive data

The data is anonymised. It nonetheless consists of recordings of human speakers, so please handle it in line with the terms of the source corpora and applicable data-protection law, and do not attempt to re-identify speakers. Access is gated and granted after manual review.

Licence

Released under CC BY-NC-SA 4.0 (Attribution-NonCommercial-ShareAlike 4.0 International), subject to the terms of the underlying ESLO and LangAge corpora. ESLO material is Copyright (c) 2012 Université d'Orléans / LLL and is made freely available for non-commercial use under a Creative Commons license; the NonCommercial term of this dataset is consistent with those terms.

Citation

If you use this dataset, please cite:

@misc{frwhisper-dataset2025,
  title={FrWhisper Training Dataset: French Conversational Speech with Interjections},
  author={Hanno Müller, Annette Gerstenberg},
  year={2025},
  note={Processed from LangAge and ESLO corpora}
}

Please also cite the ESLO and LangAge source corpora.

Authors

Funding

The AI Service Centre Berlin Brandenburg is funded by the Federal Ministry of Research, Technology and Space under the funding code 01IS22092.

Contact

For questions about this dataset, please open an issue in the dataset repository or contact kisz@hpi.de.

Downloads last month: -

Total file size:

6.09 GB