ggfox00000/stt-tedx-test-fr

---
license: cc-by-nc-nd-4.0
task_categories:
- automatic-speech-recognition
language:
- fr
size_categories:
- 1K<n<10K
pretty_name: mTEDx French — test split (utterance + long_form configs)
tags:
- mtedx
- tedx
- facebook
- french
- asr
- speech
- long-form
- public-talk
annotations_creators:
- expert-generated
source_datasets:
- extended|mtedx
dataset_info:
- config_name: utterance
features:
- name: id
dtype: string
- name: talk_id
dtype: string
- name: cue_idx
dtype: int32
- name: start_sec
dtype: float64
- name: end_sec
dtype: float64
- name: duration_sec
dtype: float64
- name: transcript
dtype: string
- name: audio
dtype:
audio:
sampling_rate: 16000
splits:
- name: test
num_examples: 2007
- config_name: long_form
features:
- name: talk_id
dtype: string
- name: audio
dtype:
audio:
sampling_rate: 16000
- name: transcript
dtype: string
- name: duration_total_sec
dtype: float64
- name: n_segments
dtype: int32
- name: segments
list:
- name: start
dtype: float64
- name: end
dtype: float64
- name: text
dtype: string
splits:
- name: test
num_examples: 10
configs:
- config_name: utterance
default: true
data_files:
- split: test
path: data/test-*.parquet
- config_name: long_form
data_files:
- split: test
path: data/long_form/test-*.parquet
---

# mTEDx FR — test split (mirror of `facebook/mtedx`)

Mirror **public** du split `test` de **mTEDx FR** (Salesky et al. 2021,
Facebook AI), corpus de talks TEDx français.

> Ce repo expose **deux configs** au choix selon ton besoin de bench :
>
> - **`utterance`** (défaut historique) — 2 007 rows courts (3-10 sec),
> découpés via les **cues VTT** TED. Idéal pour bench WER utterance-level
> (comparable à FLEURS, MLS, CoVoST-2, VoxPopuli).
> - **`long_form`** — 10 rows = 10 talks **entiers** (8-13 min chacun, total
> ~1.69 h). Idéal pour bench Whisper en **conditions réelles** :
> chunked decoding, dérive temporelle, cohérence inter-chunk.

## Configs

### `utterance` — découpage VTT (utterance-level)
- 2 007 utterances (3-10 sec, 16 kHz mono FLAC)
- Découpées via les timecodes des cues VTT (sous-titres TED).
- Filtre intro "Traducteur/Relecteur:" si match en début (<15s).
- Schéma : `id`, `start_sec`, `end_sec`, `duration_sec`, `talk_id`, `cue_idx`,
`audio`, `transcript`.
- Réf WER : `transcript`.

```python
from datasets import load_dataset
ds = load_dataset("ggfox00000/stt-tedx-test-fr", "utterance", split="test")
```

### `long_form` — talks entiers
- 10 rows = 10 talks TEDx **entiers** (8-13 min, 16 kHz mono FLAC)
- Audio resamplé bit-pour-bit du local upstream (pas de re-découpage).
- `segments` = segmentation **officielle mTEDx** (Salesky et al.) parsée
depuis `txt/segments` + `txt/test.fr` upstream — **plus rigoureuse** que
les VTT TED utilisées par la config `utterance`.
- Schéma : `talk_id`, `audio`, `transcript`, `duration_total_sec`,
`n_segments`, `segments` (list de `{start, end, text}`).
- Cas d'usage : bench long-form Whisper (mode chunked, `return_timestamps`),
comparaison long-form vs utterance, évaluation cohérence inter-chunk.

```python
from datasets import load_dataset
ds = load_dataset("ggfox00000/stt-tedx-test-fr", "long_form", split="test")
sample = ds[0]
print(sample["talk_id"], sample["duration_total_sec"], sample["n_segments"])
print(sample["audio"]["sampling_rate"], sample["audio"]["array"].shape)
print("first segment:", sample["segments"][0])
```

## Pré-traitement

### `utterance`
- Audio source : 1 FLAC stéréo par talk (44.1k ou 48k upstream).
- **Découpage** par les cues VTT (`audio[start:end]`).
- Resampling 48k stéréo → 16k mono via `numpy.mean(axis=1)` + `soxr.resample(quality="HQ")`.
- Encodé FLAC mono PCM_16.

### `long_form`
- Audio source : 1 FLAC stéréo par talk (44.1k ou 48k upstream, 8-13 min).
- **Aucun découpage temporel** — talk préservé entier.
- Resampling stéréo → mono + 44.1k/48k → 16k via numpy mean + soxr HQ.
- Encodé FLAC mono PCM_16.
- `segments` parsée depuis :
- `txt/segments` upstream (lignes `<seg_id> <talk_id> <start> <end>`)
- `txt/test.fr` upstream (1 ligne = 1 segment, ordre identique à `segments`)

## Source

- Dataset upstream : `facebook/mtedx` (https://www.openslr.org/100/)
- Paper : Salesky et al. 2021, *"The Multilingual TEDx Corpus for Speech
Recognition and Translation"* (Interspeech 2021)
- Source brute : talks TEDx publics (https://www.ted.com)

## Licence

**CC-BY-NC-ND-4.0** (héritée de mTEDx upstream — non-commercial, pas de
modifications redistribuables au-delà du resampling/encodage technique).

## Citation

```bibtex
@inproceedings{salesky2021mtedx,
title = {{The Multilingual TEDx Corpus for Speech Recognition and
Translation}},
author = {Salesky, Elizabeth and others},
booktitle = {Proceedings of Interspeech 2021},
year = {2021},
}
```