Hugging Face's logo

Datasets:
OpenLLM-France
/
Luciole-Audio-Training-Dataset

Tasks:
Automatic Speech Recognition
Translation
Question Answering
Modalities:
Text
Audio
Formats:
json
Languages:
English
French
multilingual
Size:
10M - 100M
ArXiv:
Tags:
audio
speech
music
sound
audio-language-model
speech-llm
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer (First 5GB)
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Not found.
Error code:   ResponseNotFound

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Luciole Audio Training Dataset

Dataset description

Luciole Audio Training Dataset is a large, multilingual, multi-task collection of audio–text conversations used to train the OpenLLM-France Luciole audio-language models. It adapts a text LLM to understand audio by pairing speech, music and environmental sounds with instruction-style dialogues (transcription, translation, spoken question answering, audio/music/sound captioning and question answering, speaker and language attributes, diarization, and more).

The repository holds only the text and metadata (the conversations and their references to audio files). The audio itself is not hosted on the Hub — see Getting the audio.

⚠️ Licensing / redistribution. This dataset aggregates many source datasets under different licenses. See Datasets and licenses below: every split maps to exactly one source dataset, with its license and where its audio can be obtained. Datasets whose license forbids redistribution are not included here and must be obtained from their original source.

Dataset structure 

Luciole-Audio-Training-Dataset/
├── README.md
├── dataset_stats.csv        # per-split #examples / #audio / hours / license / audio_hosted
├── dataset_index.json       # split → source dataset, license, provenance
└── data/                    # the conversations (this is what lives on the Hub)
    ├── speech/  (asr, ast, qa, summarization, diarization, temporal, emotion_reco,
    │             gender_reco, age_reco, language_reco, voice_captioning, …)
    ├── music/   (qa, captioning)
    └── sound/   (qa, captioning)

The two top levels are <domain>/<task> (e.g. speech/asr, music/qa, sound/qa). Each <domain>/<task> is exposed as a configuration <domain>.<task>, and each source dataset (per language) is a split within it. For example, English ASR from Common Voice is the CommonVoice_en split of the speech.asr config.

Data fields

Every record, in every split, has exactly these four fields:

field type description
id string unique example id.
conversations list the dialogue turns (see below).
language string "fr", "en", … ; "multilingual" when a record mixes languages; for translation, source-target like "en-fr" (English → French).
extra string a JSON-serialized object with any extra source metadata; dataset_name is present for most records. Example: {"dataset_name": "fleurs"}. "{}" when there is none.

Each turn in conversations has:

field type description
from string "User" or "Assistant"
type string "text" or "audio"
value string the text, or a relative path to an audio file (for type: "audio")
duration float audio length in seconds (audio turns only; null otherwise)
offset float start offset in the audio file, in seconds (when the turn uses a segment; null otherwise)

Audio is referenced by a relative path of the form audio/<domain>/<Dataset>/<file>, e.g. audio/speech/FLEURS_fr/16207707140941618664.wav. The path is task-independent and the <Dataset> folder carries the audio's own language when it is speech (CommonVoice_en, Multilingual_LibriSpeech_de, …) but no language when the audio has none (FMA_GenreQA for music). As a result a clip that is reused across several splits — the same recording serving both transcription and translation, or one music track with Q&A in several languages — is stored once and referenced by each split, instead of being copied per task or per text-language. (The data/ splits themselves keep their language-suffixed names.)

Example record:

{
  "id": "735",
  "conversations": [
    {"from": "User", "value": "Faites une transcription complète du fichier audio.", "type": "text"},
    {"from": "User", "value": "audio/speech/FLEURS_fr/16207707140941618664.wav", "type": "audio", "duration": 8.88},
    {"from": "Assistant", "value": "Les permis doivent être réservés à l'avance. …", "type": "text"}
  ],
  "language": "fr",
  "extra": "{\"dataset_name\": \"fleurs\"}"
}

Tasks

The dataset covers the following tasks, organised by domain as <domain>/<task> (each exposed as the configuration <domain>.<task>):

Speech

  • asr — Automatic Speech Recognition: transcribe spoken audio into text.
  • ast — Automatic Speech Translation: translate spoken audio into text in another language (the split's language is the source-target pair, e.g. en-fr for the translation from English to French).
  • qa — Spoken Question Answering: answer questions about spoken audio.
  • summarization — Summarization: produce a summary of a spoken talk or meeting (e.g. AMI/ICSI meetings, TED talks).
  • diarization — Speaker Diarization: segment the audio by speaker ("who spoke when"), with timestamps. These conversations were generated by OpenLLM-France from the speaker-turn annotations of five meeting/conversation corpora — AMI and ICSI (English), CFPP2000, ESLO and TCOF (French). The audio is re-used unchanged, segmented into short windows (~10–90 s); the assistant target comes in three flavours — a speaker-labelled transcript, a "who spoke when" timeline (speaker + start/end times), or the two combined — rendered in a variety of speaker-label and timestamp formats. Speakers are anonymized as "Speaker 1/2/3…" by order of first appearance, with window-local timestamps.
  • temporal — Temporal Localization: locate words/sentences in time (word↔time, time↔sentence) and answer questions together with the timestamp of the answer. These conversations were generated by OpenLLM-France from forced-aligned transcripts (word-level start/end times) of Multilingual LibriSpeech (French) and SLUE-SQA-5 (English): for each utterance we sample a target word or timestamp and fill question/answer templates covering the four directions — word→time, time→word, time→sentence, word→sentence — and mix in ~5% deliberately unanswerable cases (a word not spoken in the clip, or a time past its end) so the model learns to decline when there is no answer. Both short exact answers and natural-language phrasings are produced.
  • emotion_reco — Emotion Recognition: identify the emotion conveyed by the speaker.
  • gender_reco — Gender Recognition: identify the speaker's (self-reported / perceived) gender.
  • age_reco — Age Recognition: estimate the speaker's age range.
  • language_reco — Spoken Language Identification: identify which language is being spoken.
  • voice_captioning — Voice Captioning: describe the speaker's voice and acoustic characteristics (pitch, tone, accent, recording environment…).
  • sentence_stress_detection — Sentence-Stress Detection: transcribe the audio and mark which words are stressed / emphasized.
  • sentence_stress_reasoning — Sentence-Stress Reasoning: reason about how the stressed words change the meaning or intent of the utterance.
  • task_switching — Multi-turn, Multi-task Conversations: dialogues that combine several of the above tasks (and several audio clips) within a single conversation.

Music

  • qa — Music Question Answering: answer questions about a music clip (genre, instruments, mood, tempo…).
  • captioning — Music Captioning: produce a descriptive caption of a music clip.

Sound

  • qa — Environmental-sound Question Answering: answer questions about non-speech / everyday sounds.
  • captioning — Environmental-sound Captioning: describe non-speech / ambient sounds.

Loading

Because each <domain>/<task> is a separate config, load a config and pick a split:

from datasets import load_dataset

asr = load_dataset("OpenLLM-France/Luciole-Audio-Training-Dataset", "speech.asr")
fleurs_fr = load_dataset("OpenLLM-France/Luciole-Audio-Training-Dataset",
                         "speech.asr", split="FLEURS_fr")

See dataset_stats.csv for the size of every split and dataset_index.json for the exact source dataset/license behind each one.

Getting the audio

The audio is not part of this Hub repository. Most of it is served separately at:

https://dl.labs.linagora.com/files/datasets/OpenLLM-France/Luciole-Audio-Training-Dataset/audio/

Download that audio/ folder and place it next to data/; its layout mirrors data/ (audio/<domain>/<task>/<Split>/…), so each conversation's audio path resolves directly. The folder can be downloaded in zip format (tar, targz and tarbz2 also supported) with the following command:

TOKEN=$(curl -s https://dl.labs.linagora.com/api/login \
  -H "Content-Type: application/json" -d '{}')

curl -H "X-Auth: $TOKEN" \
  "https://dl.labs.linagora.com/api/raw/datasets/OpenLLM-France/Luciole-Audio-Training-Dataset/audio/?algo=zip" \
  -o <output-name>.zip

Some datasets' audio is not redistributable by us (copyrighted source, YouTube-only, etc.). For those, the conversations are still published but you must download the audio from the original source listed below. See Audio availability.

Datasets and licenses

This dataset is an aggregation of third-party datasets, each under its own license. The table below lists every source dataset used, its license, what is published here, where to get the audio, and its content (total samples and audio duration, broken down per task and language).

The instruction prompts (the user-turn wording that states each task) were written by OpenLLM-France for every task except question answering (qa).

Marks in the Content column flag data produced by OpenLLM-France, on the specific task line(s) concerned: 🎙️ = audio synthesized by OpenLLM-France · ✍️ = text (questions/answers, translations, or captions) generated or modified by OpenLLM-France — see each dataset's note below for details.

Dataset License In this release Audio Content
ACSYNT CC-BY-SA-4.0 ✓ text + audio hosted 3,539 samples · 7h54 — asr, fr
African Accented French Apache-2.0 ✓ text + audio hosted 11,470 samples · 14h — asr, fr
AMI Meeting Corpus CC-BY-4.0 ✓ text + audio hosted 10,343 samples · 183h
• ✍️ diarization, en — 9,557 · 113h
• summarization, en — 786 · 70h
amuvarma/10k-filtered-tune-audio Not declared ✗ excluded (external only) source 9,000 samples · 13h — qa, en
Arabic Speech Corpus (Halabi) CC-BY-4.0 ✓ text + audio hosted 1,813 samples · 3h49 — asr, ar
AudioCaps MIT (text)
Copyrighted (audio)		 ✓ text only source 90,356 samples · 248h
• sound.captioning, en — 45,178 · 124h
• ✍️ sound.captioning, fr — 45,178 · 124h
aya_collection Apache-2.0 (text)
MIT (audio)		 ✓ text + audio hosted 723 samples · 1h — 🎙️ qa, fr
CFPB CC-BY-NC-SA-4.0 ✓ text + audio hosted 18,415 samples · 9h05 — asr, fr
CFPP2000 CC-BY-NC-SA-3.0 ✓ text + audio hosted 35,786 samples · 93h
• asr, fr — 30,278 · 38h
• ✍️ diarization, fr — 5,508 · 55h
CLAPI CC-BY-NC-SA-4.0 ✓ text only source 17,799 samples · 22h — asr, fr
CLEAR CC-BY-3.0 (text)
CC-BY-NC-4.0 (audio)		 ✓ text + audio hosted 1,425,051 samples · 17603h
• sound.qa, en — 712,527 · 8801h
• sound.qa, fr — 712,524 · 8801h
Clotho (via AudioMCQ) Tampere University NC (text)
Mixed CC per-file (audio)		 ✓ text + audio hosted 3,829 samples · 24h
• sound.qa, en — 3,801 · 24h
• sound.qa, fr — 28 · 10min
Clotho-AQA MIT (text)
Mixed CC per-file (audio)		 ✓ text + audio hosted 13,527 samples · 85h
• ✍️ sound.qa, en — 5,802 · 36h
• ✍️ sound.qa, fr — 5,802 · 36h
• task_switching, en — 1,923 · 12h
Common Voice CC0-1.0 ✓ text + audio hosted 11,040,447 samples · 16492h
• asr, en — 1,138,760 · 1798h
• ✍️ ast, en-fr — 1,138,760 · 1798h
• ✍️ gender_reco, en — 780,521 · 1228h
• ✍️ age_reco, en — 667,207 · 1036h
• asr, de — 607,871 · 963h
• ✍️ ast, de-fr — 607,530 · 962h
• asr, fr — 593,066 · 851h
• ✍️ ast, fr-en — 593,033 · 851h
• ✍️ ast, fr-it — 592,957 · 851h
• ✍️ ast, fr-pt — 592,952 · 851h
• ✍️ ast, fr-es — 592,951 · 851h
• ✍️ ast, fr-nl — 592,709 · 850h
• ✍️ ast, fr-de — 592,511 · 850h
• ✍️ gender_reco, fr — 365,902 · 520h
• asr, es — 353,701 · 509h
• ✍️ ast, es-fr — 353,539 · 508h
• ✍️ age_reco, fr — 288,112 · 408h
• asr, it — 172,828 · 255h
• ✍️ ast, fr-ar — 116,027 · 166h
• ✍️ ast, it-fr — 109,725 · 162h
• asr, nl — 43,458 · 54h
• ✍️ ast, nl-fr — 43,448 · 54h
• asr, ar — 28,531 · 32h
• ✍️ ast, ar-fr — 28,508 · 32h
• asr, pt — 22,923 · 26h
• ✍️ ast, pt-fr — 22,917 · 26h
CommonVoice Multi-Turn CC0-1.0 ✓ text + audio hosted 1,000,000 samples · 1492h — ✍️ task_switching, multilingual
CompA-R (via AudioMCQ) Apache-2.0 (text)
Copyrighted (audio)		 ✓ text only source 200,929 samples · 552h
• sound.qa, en — 197,218 · 542h
• sound.qa, fr — 3,711 · 10h
Compar:IA Etalab-2.0 (text)
MIT (audio)		 ✓ text + audio hosted 12,900 samples · 21h — 🎙️ qa, fr
CoVoST / CoVoST2 CC0 (text)
CC0-1.0 (audio)		 ✓ text + audio hosted 1,519,941 samples · 2046h
• ast, en-ar — 429,267 · 588h
• ast, en-de — 429,267 · 588h
• ast, de-en — 264,140 · 346h
• ast, fr-en — 247,231 · 315h
• ast, es-en — 91,902 · 131h
• ast, it-en — 42,867 · 61h
• ast, pt-en — 14,894 · 18h
• ast, ar-en — 373 · 22min
emotional-speech-audio-dataset-4languages CC-BY-NC-SA-4.0 ✓ text + audio hosted 144 samples · 10min — ✍️ emotion_reco, fr
ESCWA CC-BY-NC-4.0 (text)
Copyrighted (audio)		 ✗ excluded (external only) source 845 samples · 2h47 — asr, ar
ESLO CC-BY-NC-SA-4.0 ✓ text + audio hosted 498,087 samples · 677h
• asr, fr — 454,574 · 272h
• ✍️ diarization, fr — 43,513 · 404h
FCaps CC-BY-NC-SA-4.0 (text)
CC-BY-NC-4.0 + Copyrighted (audio)		 ✓ text only source 109,790 samples · 256h — voice_captioning, en
FLEURS CC-BY-4.0 ✓ text + audio hosted 22,404 samples · 68h
• asr, fr — 3,193 · 10h
• asr, it — 3,028 · 8h59
• asr, de — 2,973 · 8h58
• asr, nl — 2,918 · 7h39
• asr, es — 2,796 · 8h48
• asr, pt — 2,790 · 10h
• asr, en — 2,602 · 7h29
• asr, ar — 2,104 · 6h02
FMA (FMA_GenreQA) CC-BY-4.0 (text)
CC-BY / CC0 / Public Domain (audio)		 ✓ text + audio hosted 12,616 samples · 101h
• ✍️ music.qa, en — 6,308 · 51h
• ✍️ music.qa, fr — 6,308 · 51h
gruhit-patel/alpaca_speech_instruct CC-BY-NC-4.0 ✓ text + audio hosted 51,753 samples · 72h — qa, en
HeySQuAD CC-BY-SA-4.0 (text)
CC-BY-4.0 (audio)		 ✓ text + audio hosted 10,757 samples · 18h — qa, en
ICSI Meeting Corpus CC-BY-4.0 ✓ text + audio hosted 9,554 samples · 112h
• ✍️ diarization, en — 9,144 · 89h
• summarization, en — 410 · 23h
JamendoMaxCaps CC-BY-SA-3.0 (text)
Mixed CC (audio)		 ✓ text only source 721,497 samples · 5764h
• ✍️ music.captioning, en — 360,764 · 2882h
• ✍️ music.captioning, fr — 360,733 · 2882h
LesVocaux CC-BY-NC-SA-4.0 ✓ text + audio hosted 744 samples · 10h — asr, fr
LeVoiceLab CTFAR/CTFNN1 Private ✗ excluded (external only) source 127,990 samples · 125h — asr, fr
liangtianle/science-question Not declared ✗ excluded (external only) source 3,500 samples · 7h43 — qa, en
liangtianle/unsafety-question Not declared ✗ excluded (external only) source 3,383 samples · 16h — qa, en
LINAGORA Meetings CC-BY-SA-4.0 ✓ text + audio hosted 6,197 samples · 10h — asr, fr
LP-MusicCaps-MTT (via AudioMCQ) CC-BY-NC-4.0 (text)
CC-BY-NC-SA-3.0 (audio)		 ✓ text + audio hosted 15,754 samples · 127h
• music.qa, en — 15,560 · 126h
• music.qa, fr — 194 · 1h34
MASC CC-BY-4.0 (text)
Copyrighted (audio)		 ✓ text only source 413,499 samples · 544h — asr, ar
MECAT-Caption CC-BY-3.0 (text)
Copyrighted (audio)		 ✓ text only source 39,664 samples · 111h
• ✍️ sound.captioning, en — 19,832 · 55h
• ✍️ sound.captioning, fr — 19,832 · 55h
MECAT-QA CC-BY-3.0 (text)
Copyrighted (audio)		 ✓ text only source 200,390 samples · 559h
• sound.qa, en — 100,195 · 280h
• sound.qa, fr — 100,195 · 280h
MELD GPL-3.0 (text)
Copyrighted (audio)		 ✓ text only source 24,557 samples · 21h
• ✍️ emotion_reco, en — 9,988 · 8h43
• ✍️ emotion_reco, fr — 9,988 · 8h43
• voice_captioning, en — 4,581 · 3h59
Menlo/instruction-speech-encodec-v1.5 MIT ✓ text + audio hosted 332,367 samples · 678h — qa, en
MGB-2 Research-only EULA (text)
Copyrighted (audio)		 ✗ excluded (external only) source 274,073 samples · 825h — asr, ar
MGB-5 Unknown (text)
Copyrighted (audio)		 ✗ excluded (external only) source 1,022 samples · 1h54 — asr, ar
MiscMultiTurn Mixed ✗ excluded (external only) source 99,980 samples · 428h — task_switching, multilingual
MLS Multi-Turn CC-BY-4.0 ✓ text + audio hosted 73,308 samples · 810h — ✍️ task_switching, fr
Multilingual LibriSpeech CC-BY-4.0 ✓ text + audio hosted 12,545,420 samples · 54609h
• asr, en — 10,808,037 · 44660h
• asr, de — 469,942 · 1967h
• asr, nl — 374,287 · 1554h
• ✍️ temporal, fr — 317,084 · 4026h
• asr, fr — 258,213 · 1077h
• asr, es — 220,701 · 918h
• asr, it — 59,623 · 247h
• asr, pt — 37,533 · 161h
Multilingual TEDx CC-BY-NC-ND-4.0 ✗ excluded (external only) source 577,116 samples · 975h
• asr, fr — 116,045 · 176h
• asr, es — 102,171 · 178h
• asr, pt — 90,244 · 153h
• asr, it — 49,964 · 101h
• ast, es-en — 36,263 · 64h
• ast, pt-en — 30,855 · 53h
• ast, fr-en — 30,171 · 45h
• ast, it-en — 24,576 · 48h
• ast, es-pt — 21,107 · 37h
• ast, fr-es — 20,826 · 33h
• ast, fr-pt — 13,286 · 20h
• asr, ar — 11,821 · 16h
• ast, pt-es — 11,499 · 21h
• asr, de — 6,764 · 10h
• ast, es-it — 5,600 · 9h55
• ast, es-fr — 3,663 · 6h20
• ast, it-es — 2,261 · 4h44
MusicCaps CC-BY-SA-4.0 (text)
Copyrighted (audio)		 ✓ text only source 10,320 samples · 29h
• music.captioning, en — 2,580 · 7h10
• ✍️ music.captioning, fr — 2,580 · 7h10
• ✍️ music.qa, en — 2,580 · 7h10
• ✍️ music.qa, fr — 2,580 · 7h10
MusicCaps (via AudioMCQ) CC-BY-SA-4.0 (text)
Copyrighted (audio)		 ✓ text only source 2,606 samples · 7h14
• music.qa, en — 2,565 · 7h07
• music.qa, fr — 41 · 6min
MusicQA MIT (text)
Copyrighted (audio)		 ✓ text only source 26,204 samples · 227h
• music.qa, en — 13,102 · 114h
• music.qa, fr — 13,102 · 114h
Nemotron Post-Training v3 CC-BY-4.0 (text)
N/A (audio)		 ✗ excluded (external only) source 289,784 samples — text_only, en
Nutshell CC-BY-4.0 ✓ text + audio hosted 3,981 samples · 808h — summarization, en
OreauFR_02 CC-BY-4.0 ✓ text + audio hosted 434 samples · 20min — ✍️ emotion_reco, fr
PFC CC-BY-NC-4.0 ✓ text + audio hosted 67,374 samples · 62h — asr, fr
PxSLU / PxCorpus CC-BY-4.0 ✓ text + audio hosted 1,955 samples · 4h12 — asr, fr
SIFT-50M (Common Voice subset) CDLA-Sharing-1.0 (text)
CC0-1.0 (audio)		 ✓ text + audio hosted 1,932,205 samples · 2948h
• qa, de — 710,082 · 1135h
• qa, fr — 521,557 · 754h
• gender_reco, de — 232,062 · 366h
• qa, it — 170,900 · 253h
• gender_reco, it — 117,585 · 170h
• age_reco, it — 99,704 · 145h
• age_reco, de — 42,523 · 71h
• qa, es — 20,061 · 29h
• gender_reco, es — 8,955 · 13h
• age_reco, es — 8,776 · 13h
SIFT-50M (MLS subset) CDLA-Sharing-1.0 (text)
CC-BY-4.0 (audio)		 ✓ text + audio hosted 1,897,963 samples · 7936h
• qa, de — 621,424 · 2621h
• qa, fr — 407,276 · 1713h
• gender_reco, de — 319,987 · 1320h
• qa, es — 297,437 · 1248h
• gender_reco, es — 132,216 · 537h
• qa, it — 77,151 · 323h
• gender_reco, it — 42,472 · 174h
SimSamu MIT ✓ text + audio hosted 3,089 samples · 2h36 — asr, fr
SLUE Multi-Turn Apache-2.0 / CC-BY-SA-4.0 ✓ text + audio hosted 13,325 samples · 144h — ✍️ task_switching, en
SLUE-SQA-5 (SLUE phase-2) Apache-2.0 / CC-BY-SA-4.0 ✓ text + audio hosted 444,526 samples · 4057h
• ✍️ temporal, en — 305,968 · 3392h
• qa, en — 138,558 · 665h
SLUE-TED (SLUE phase-2) CC-BY-NC-ND-4.0 ✗ excluded (external only) source 3,384 samples · 665h — summarization, en
SpeechCraft (via AudioMCQ) Apache-2.0 (text)
CC-BY-4.0 (audio)		 ✓ text + audio hosted 21,683 samples · 47h — voice_captioning, en
ssi-speech CC-BY-NC-SA-4.0 ✓ text only source 60,000 samples · 44h
• ✍️ age_reco, en — 10,000 · 7h17
• ✍️ age_reco, fr — 10,000 · 7h17
• ✍️ emotion_reco, en — 10,000 · 7h17
• ✍️ emotion_reco, fr — 10,000 · 7h17
• ✍️ gender_reco, en — 10,000 · 7h17
• ✍️ gender_reco, fr — 10,000 · 7h17
Stress-17K CC-BY-NC-4.0 ✓ text + audio hosted 13,200 samples · 11h
• sentence_stress_detection, en — 4,400 · 3h38
• sentence_stress_detection, fr — 4,400 · 3h38
• sentence_stress_reasoning, en — 4,400 · 3h38
TACOS (via AudioMCQ) Apache-2.0 (text)
Mixed Freesound per-file (audio)		 ✓ text + audio hosted 34,506 samples · 213h
• sound.qa, en — 33,320 · 206h
• sound.qa, fr — 1,186 · 7h20
TARIC CC-BY-NC-4.0 ✓ text + audio hosted 15,034 samples · 6h55 — asr, ar
TCOF CC-BY-NC-SA-2.0-FR ✓ text + audio hosted 171,212 samples · 272h
• asr, fr — 152,737 · 108h
• ✍️ diarization, fr — 18,475 · 165h
TunSwitch CC-BY-4.0 (text)
Broadcast-derived (audio)		 ✓ text only source 5,321 samples · 8h17 — asr, ar
UltraChat-300K-SLAM-Omni MIT ✓ text + audio hosted 164,464 samples · 342h — qa, en
Vigogne (French Alpaca) CC-BY-NC-4.0 (text)
MIT (audio)		 ✓ text + audio hosted 49,657 samples · 80h — 🎙️ qa, fr
VoxCeleb CC-BY-SA-4.0 (text)
Copyrighted (audio)		 ✓ text only source 200,000 samples · 402h — diarization, multilingual
VoxForge GPL-3.0-or-later ✓ text + audio hosted 22,429 samples · 37h — asr, fr
VoxLingua107 CC-BY-NC-4.0 (text)
Copyrighted (audio)		 ✓ text only source 27,000 samples · 73h
• ✍️ language_reco, en — 13,500 · 36h
• ✍️ language_reco, fr — 13,500 · 37h
VoxPopuli CC0-1.0 ✓ text + audio hosted 1,595,912 samples · 10347h
• ✍️ qa, fr — 568,923 · 4538h
• ✍️ qa, en — 568,007 · 4540h
• asr, en — 182,482 · 523h
• asr, de — 108,473 · 265h
• asr, fr — 73,561 · 206h
• asr, es — 50,922 · 152h
• asr, it — 22,576 · 78h
• asr, nl — 20,968 · 46h
WavCaps CC-BY-4.0 (text)
Mixed (audio)		 ✓ text only source 777,564 samples · 14825h
• sound.captioning, en — 388,782 · 7413h
• ✍️ sound.captioning, fr — 388,782 · 7413h
YODAS CC-BY-3.0 ✓ text only source 8,316,967 samples · 21821h
• asr, en — 7,984,260 · 19447h
• asr, fr — 332,707 · 2373h
YouTubeFr Copyrighted ✗ excluded (external only) source 1,532,799 samples · 5250h — asr, fr

Audio availability

Where to obtain the audio for each source dataset, grouped by what is published in this repository.

Text and audio

Both the conversations and the audio are published. The audio is hosted at https://dl.labs.linagora.com/files/datasets/OpenLLM-France/Luciole-Audio-Training-Dataset/audio/; each record's audio path is relative to that folder.

Text only

The conversations are published here, but the audio is not redistributable by us — download it from the original source(s):

For YouTube-sourced audio (e.g. MusicCaps, YouTubeFr): the original video/clip ID is preserved in the audio file name (and in each record's extra field). Fetch the audio from YouTube with the source dataset's tooling or a downloader such as yt-dlp, then cut each segment using the offset / duration of the corresponding audio turn.

Not included

Neither the audio nor the conversations are published (the license forbids redistribution) — get the whole dataset from the source(s):

Licensing

No single license covers the whole collection — it is a collection in which each source dataset retains its own license (see the table above). The curator's added layer (the instruction wrappers, curation and formatting) and the collection as a whole are offered under CC-BY-NC-SA-4.0, which is the most permissive single license consistent with the NonCommercial + ShareAlike terms of the included sources:

  • NonCommercial (NC): several sources are NC, so the combined dataset is non-commercial.
  • ShareAlike (SA): several sources are NC-SA, so adaptations must stay under the same license.
  • Attribution (BY): attribute the original datasets (this table) when you use it.

A few sources are CC-BY-SA (without NC); they are kept as separate splits (a collection, not a remix) so their ShareAlike terms are respected without relicensing.

NoDerivatives (ND) datasets (e.g. Multilingual TEDx, SLUE-TED) may not be redistributed in modified form. Our pipeline re-segments and reformats audio into conversations, which is a derivative, so these are excluded — download them from their original source (see the table).

When in doubt, comply with the most restrictive applicable terms for the subset you use. If you are a rights holder and believe something should not be included, please open an issue.

Attribution

This dataset builds on the work of many others. Several sources are under attribution (CC-BY)* licenses and require credit. Please attribute every source dataset you use (the table above lists all of them with their license and origin); the attribution-required ones are highlighted here:

The following datasets carry a CC-BY / attribution license on their text and/or audio — you must credit them (and cite their papers) when you use this dataset:

  • ACSYNT (CC-BY-SA-4.0) — source
  • AMI Meeting Corpus (CC-BY-4.0) — source
  • Arabic Speech Corpus (Halabi) (CC-BY-4.0) — source
  • CFPB (CC-BY-NC-SA-4.0) — source
  • CFPP2000 (CC-BY-NC-SA-3.0) — source
  • CLAPI (CC-BY-NC-SA-4.0) — source
  • CLEAR (CC-BY-3.0 (text), CC-BY-NC-4.0 (Good-Sounds / Zenodo) (audio)) — source
  • Clotho (via AudioMCQ) (Tampere University NC (text), Mixed CC per-file (CC0 … CC-BY-NC-3.0; Freesound) (audio)) — source
  • Clotho-AQA (MIT (text), Mixed CC per-file (CC0 … CC-BY-NC-3.0; Freesound) (audio)) — source
  • emotional-speech-audio-dataset-4languages (CC-BY-NC-SA-4.0) — source
  • ESLO (CC-BY-NC-SA-4.0) — source
  • FCaps (CC-BY-NC-SA-4.0 (text), CC-BY-NC-4.0 (EARS/Expresso) + Copyrighted (VoxCeleb) (audio)) — source
  • FLEURS (CC-BY-4.0) — source
  • FMA (FMA_GenreQA) (CC-BY-4.0 (text), CC-BY / CC0 / Public Domain (audio)) — source
  • gruhit-patel/alpaca_speech_instruct (CC-BY-NC-4.0) — source
  • HeySQuAD (CC-BY-SA-4.0 (SQuAD 1.1) (text), CC-BY-4.0 (audio)) — source
  • ICSI Meeting Corpus (CC-BY-4.0) — source
  • JamendoMaxCaps (CC-BY-SA-3.0 (text), Mixed CC (Jamendo per-track, incl. CC-BY-NC-ND-3.0) (audio)) — source
  • LesVocaux (CC-BY-NC-SA-4.0) — source
  • LINAGORA Meetings (CC-BY-SA-4.0) — source
  • LP-MusicCaps-MTT (via AudioMCQ) (CC-BY-NC-4.0 (text), CC-BY-NC-SA-3.0 (MagnaTagATune) (audio)) — source
  • MASC (CC-BY-4.0 (text), Copyrighted (YouTube) (audio)) — source
  • MECAT-Caption (CC-BY-3.0 (text), Copyrighted (YouTube / ACAV100M) (audio)) — source
  • MECAT-QA (CC-BY-3.0 (text), Copyrighted (YouTube / ACAV100M) (audio)) — source
  • MLS Multi-Turn (CC-BY-4.0) — source
  • Multilingual LibriSpeech (CC-BY-4.0) — source
  • MusicCaps (CC-BY-SA-4.0 (text), Copyrighted (YouTube / music industry rights) (audio)) — source
  • MusicCaps (via AudioMCQ) (CC-BY-SA-4.0 (text), Copyrighted (YouTube / AudioSet) (audio)) — source
  • Nutshell (CC-BY-4.0) — source
  • OreauFR_02 (CC-BY-4.0) — source
  • PFC (CC-BY-NC-4.0) — source
  • PxSLU / PxCorpus (CC-BY-4.0) — source
  • SIFT-50M (MLS subset) (CDLA-Sharing-1.0 (text), CC-BY-4.0 (Multilingual LibriSpeech) (audio)) — source
  • SLUE Multi-Turn (Apache-2.0 / CC-BY-SA-4.0) — source
  • SLUE-SQA-5 (SLUE phase-2) (Apache-2.0 / CC-BY-SA-4.0) — source
  • SpeechCraft (via AudioMCQ) (Apache-2.0 (text), CC-BY-4.0 (LibriTTS-R) (audio)) — source
  • ssi-speech (CC-BY-NC-SA-4.0) — source
  • Stress-17K (CC-BY-NC-4.0) — source
  • TACOS (via AudioMCQ) (Apache-2.0 (text), Mixed Freesound per-file (CC0 … CC-BY-NC-4.0) (audio)) — source
  • TARIC (CC-BY-NC-4.0) — source
  • TCOF (CC-BY-NC-SA-2.0-FR) — source
  • TunSwitch (CC-BY-4.0 (text), Broadcast-derived (radio/TV) — not redistributable (audio)) — source
  • Vigogne (French Alpaca) (CC-BY-NC-4.0 (text), MIT (TTS-synthesized) (audio)) — source
  • VoxCeleb (CC-BY-SA-4.0 (text), Copyrighted (YouTube-sourced) (audio)) — source
  • VoxLingua107 (CC-BY-NC-4.0 (text), Copyrighted (YouTube) (audio)) — source
  • WavCaps (CC-BY-4.0 (annotations; card states academic-use-only) (text), Mixed (FreeSound per-clip CC / BBC / SoundBible / AudioSet) — not redistributable (audio)) — source
  • YODAS (CC-BY-3.0) — source

These also require keeping their text license/attribution notice (Apache-2.0 / MIT / CDLA):

  • African Accented French (Apache-2.0) — source
  • AudioCaps (MIT (annotations; README also states academic-use-only)) — source
  • aya_collection (Apache-2.0) — source
  • Clotho-AQA (MIT) — source
  • CompA-R (via AudioMCQ) (Apache-2.0) — source
  • Menlo/instruction-speech-encodec-v1.5 (MIT) — source
  • MusicQA (MIT) — source
  • SIFT-50M (Common Voice subset) (CDLA-Sharing-1.0) — source
  • SIFT-50M (MLS subset) (CDLA-Sharing-1.0) — source
  • SimSamu (MIT) — source
  • SLUE Multi-Turn (Apache-2.0 / CC-BY-SA-4.0) — source
  • SLUE-SQA-5 (SLUE phase-2) (Apache-2.0 / CC-BY-SA-4.0) — source
  • SpeechCraft (via AudioMCQ) (Apache-2.0) — source
  • TACOS (via AudioMCQ) (Apache-2.0) — source
  • UltraChat-300K-SLAM-Omni (MIT) — source

Citation

If you use this dataset, please cite the OpenLLM-France / Luciole project and the individual source datasets you rely on (see the table above for sources).

Acknowledgements

Built by LINAGORA / OpenLLM-France. This collection would not exist without the authors of the many source datasets it builds upon.
Downloads last month
-
Number of rows:
45,236,540
Estimated number of rows:
42,758,860
Total file size:
26.4 GB

Models trained or fine-tuned on OpenLLM-France/Luciole-Audio-Training-Dataset

Collection including OpenLLM-France/Luciole-Audio-Training-Dataset

Papers for OpenLLM-France/Luciole-Audio-Training-Dataset