The dataset viewer is not available for this split.
Error code: TooBigContentError
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.
SUMM-RE — French test split (mirror of linagora/SUMM-RE)
Mirror public du split test de SUMM-RE (LINAGORA / Aix-Marseille
LPL), pour benchmark ASR français conversationnel (parole de réunion,
3-4 locuteurs, ~20 min par session).
⚠ Ce repo ne contient que le split test (124 tracks individuelles = 37 réunions × 3-4 micros). Pour les splits
train/dev, voir le repo upstreamlinagora/SUMM-RE.
Contenu
- 124 pistes audio individuelles (1 piste = 1 microphone d'un locuteur pour une réunion entière de ~20 min)
- 37 réunions distinctes (chacune avec 3-4 micros)
- Audio : WAV 16 kHz mono PCM_16 embarqué dans parquet (tel que upstream)
- Annotations : transcription manuelle + alignement forcé au mot
- Langue : français (fr), parole conversationnelle / réunion
- Licence : CC-BY-SA-4.0 (héritée de SUMM-RE upstream)
- Durée totale : ~41 h (124 × ~20 min)
Colonnes
| Colonne | Type | Description |
|---|---|---|
audio |
Audio | dict {{path, array, sampling_rate}} — décodé auto |
meeting_id |
string | ID réunion : <num><letter>_<scenario><type><loc> (ex 001a_PARL) |
speaker_id |
string | identifiant locuteur |
audio_id |
string | meeting_id + speaker_id (clé unique) |
transcript |
string | transcription concaténée de tous les segments (ref WER) |
segments |
list[struct] | segments avec start, end, transcript, words[start,end,word] |
Utilisation
Chargement
from datasets import load_dataset
ds = load_dataset("ggfox00000/stt-summre-fr-test", split="test")
print(ds)
sample = ds[0]
print(sample["meeting_id"], sample["speaker_id"])
print(sample["audio"]["sampling_rate"], sample["audio"]["array"].shape)
print("segments:", len(sample["segments"]))
print("first seg:", sample["segments"][0])
Découpage par segment pour évaluation WER
SUMM-RE livre 1 ligne par piste audio complète (~20 min). Pour comparer à un benchmark utterance-level, tu peux soit :
- Long-form WER : passer la track complète à Whisper en mode
chunked (
chunk_length_s=30,return_timestamps=True), comparer le résultat au champtranscript. - Utterance-level : découper par
segments[i].start/end, transcrire chaque segment isolément, comparer ausegments[i].transcript.
Évaluation diarisation
En fusionnant toutes les tracks d'une même meeting_id, tu peux évaluer
de la diarisation (qui parle quand). Les timestamps segments[i].start/end
servent de RTTM de référence. Voir paper Hunter et al. (2024) §5.
Pré-traitement
Aucun. Les 28 parquets sont transférés bit-à-bit depuis linagora/SUMM-RE
(uniquement les fichiers du split test, l'arbo data/test/test-*.parquet
upstream est aplatie en data/test-*.parquet pour cohérence avec nos autres
STT datasets ggfox00000/stt-*). Aucun resampling, aucun ré-encodage.
Source
- Dataset upstream : https://huggingface.co/datasets/linagora/SUMM-RE
- Paper : Hunter et al. 2024, "SUMM-RE: A corpus of French meeting-style conversations" (LREC-COLING 2024)
- Projet : SUMM-RE (ANR-20-CE23-0017), LINAGORA + Aix-Marseille LPL
Licence
CC-BY-SA-4.0 (héritée de SUMM-RE upstream). Toute redistribution doit préserver la licence et créditer LINAGORA + le LPL.
Citation
@inproceedings{{hunter2024summre,
title = {{SUMM-RE: A Corpus of French Meeting-Style Conversations}},
author = {{Hunter, Julie and others}},
booktitle = {{Proceedings of LREC-COLING 2024}},
year = {{2024}},
}}
Pour le pipeline d'alignement Whisper utilisé sur le split train (non
inclus dans ce mirror) :
@inproceedings{{yamasaki2023transcribing,
title = {{Transcribing And Aligning Conversational Speech: A Hybrid
Pipeline Applied To French Conversations}},
author = {{Yamasaki, Hugo and others}},
booktitle = {{Proceedings of ASRU 2023}},
year = {{2023}},
}}
- Downloads last month
- 58