Gilbert-AI/french-education-speech

French Education Speech Corpus (Phase 1)

Summary

Domain-focused French speech dataset for ASR in educational contexts. Phase 1 provides two variants:

• IPU speech-only (recommended): segments from Inter-Pausal Units (high speech density)
• VAD baseline: automatic speech segmentation (universal, slightly noisier)

Key facts

• Language: fr (French)
• Sampling: WAV 16 kHz mono
• Max segment length: ≤ 30 s
• Structure:
  • audio/ipu_conferences/*.wav (IPU speech-only)
  • audio/source_copy/_seg.wav (VAD baseline)
  • transcripts_ipu/{train.jsonl,valid.jsonl}
  • transcripts/{train.jsonl,valid.jsonl}
  • metadata.json

Intended use

• Fine-tuning and evaluation of French ASR models (e.g., Whisper) in education contexts (courses, lectures, public academic talks).

Data sources

• Phase 1 core: SUMM-RE (locally provided by user) transformed into:
  • IPU speech-only segments (recommended)
  • VAD segments (baseline)
• Additional open-access sources (pilot): individual Canal-U pages (Creative Commons). Future versions will add more CC-BY/CC-BY-SA/Etalab sources (e.g., Canal-U, MESR open videos) for category balance.

Licensing

• Phase 1 includes locally provided audio from SUMM-RE. Redistribution depends on source rights. If redistribution is restricted, distribute “metadata-only + scripts” and require local reconstruction.
• Open sources to be added use one of: CC-BY, CC-BY-SA, Etalab, or public domain. See licenses.md for per-source notes.

Schema (JSONL)

Each line in train.jsonl/valid.jsonl:

{
  "id": "string",
  "audio": "/abs/path/to/audio.wav",
  "duration": 12.34,
  "text": "transcription if available, else empty",
  "source": "SUMM-RE|Canal-U|...",
  "speaker_role": "teacher|student|researcher|admin",
  "accent": "",
  "domain": "education|science|law|general|...",
  "quality": "clean|medium|noisy",
  "category": "conferences|podcasts|meetings|interviews"
}

Stats (Phase 1)

• Total: 13,711 segments (train 12,988 / valid 723)
• Total duration: 16.57 hours
• Average segment: 4.35 seconds (median: 1.80s)
• Distribution: 67.4% conferences, 30.0% podcasts, 2.6% courses, 0.04% interviews
• Sources: SUMM-RE (IPU speech-only), Canal-U (CC-BY/CC-BY-SA)

Pipeline summary

┌─────────────────────────────────────────────────────────────────────────┐
│                         DATA COLLECTION PHASE                            │
└─────────────────────────────────────────────────────────────────────────┘

    sources.csv (URLs + metadata)
            │
            ▼
    ┌──────────────────┐
    │  download.py      │  Filter: CC-BY/CC-BY-SA/Etalab only
    │  (yt-dlp)         │  Output: audio/{category}/*.wav
    └─────────┬─────────┘
              │
              ▼
    ┌──────────────────┐
    │  normalize.sh     │  Convert to 16kHz mono WAV
    │  (ffmpeg)        │  Apply loudnorm
    └─────────┬─────────┘
              │
              ▼
    ┌──────────────────┐
    │  segment_vad.py  │  Voice Activity Detection
    │  (webrtcvad)     │  Split into chunks ≤30s
    └─────────┬─────────┘
              │
              ▼
    ┌──────────────────┐
    │ build_metadata.py│  Generate metadata.json
    │                  │  Deterministic train/valid split (hash-based)
    └─────────┬─────────┘
              │
              ▼
    ┌──────────────────┐
    │  train.jsonl     │  12,988 segments (95%)
    │  valid.jsonl     │   723 segments (5%)
    └──────────────────┘

1. Normalize to 16 kHz mono
2. Segment (IPU speech-only or VAD) with hard cap 30 s
3. Build metadata + deterministic train/valid splits (hash-based)

Reproducibility

• All scripts provided under scripts/ (normalize.sh, segment_vad.py, import_summre.py, build_metadata.py, check_balance.py, transcribe_generic.py, merge_transcripts.py, clean_texts.py).
• See docs/phase1.md and docs/replication.md for exact commands.

Quality notes

• IPU variant: high speech density, minimal silence, ideal for education content.
• VAD variant: universal, can include brief non-speech artifacts.
• All segments ≤ 30 s, WAV 16 kHz mono, deterministic split, provenance tracked.

Roadmap

• Add open, commercial-OK sources (CC-BY/CC-BY-SA/Etalab) to reach target balance:
  • 40% courses/conferences, 20% educational podcasts, 20% institutional speeches, 10% interviews, 10% Q/A.
• Provide optional baseline transcripts via a public ASR (outside this release).

Citation

Please cite this dataset card and the original data sources you use (see licenses.md).