Replication guide (Phase 1)
Environment
cd /Users/mathisescriva/Dataset-Education/FrenchEducationSpeech
python3 -m pip install -U yt-dlp tqdm pandas numpy soundfile webrtcvad pydub librosa orjson xxhash
IPU speech-only (recommended)
python scripts/import_summre.py \
--summre-root "/Users/mathisescriva/Downloads/summ-re-asru/1/SUMM-RE-sm" \
--out-audio "audio/ipu_conferences" \
--out-transcripts "transcripts_ipu" \
--train-ratio 0.95
Output: audio/ipu_conferences/*.wav, transcripts_ipu/train.jsonl, valid.jsonl
VAD baseline (optional)
mkdir -p audio/source_copy
cp /Users/mathisescriva/Downloads/summ-re-asru/1/SUMM-RE-sm/audio_anonymized/*.wav audio/source_copy/
bash scripts/normalize.sh audio/source_copy
python scripts/segment_vad.py --indir audio/source_copy --outdir audio/source_copy --max-seconds 30
python scripts/build_metadata.py --audioroot audio/source_copy --outdir transcripts --train-ratio 0.95
Open sources ingestion (per URL with open license)
- Add to sources.csv:
url,category,domain,expected_role,license,license_url,notes
https://.../video,conferences,education,teacher,CC-BY,https://...,Page shows CC-BY
- Process:
python scripts/download.py --catalog sources.csv --outdir audio
bash scripts/normalize.sh audio
python scripts/segment_vad.py --indir audio --outdir audio --max-seconds 30
python scripts/build_metadata.py --audioroot audio --outdir transcripts --train-ratio 0.95
Balance check
python scripts/check_balance.py --jsonl transcripts/train.jsonl --targets conferences=0.4 podcasts=0.2 meetings=0.2 interviews=0.1
Merging splits (example)
cat transcripts/train.jsonl transcripts_ipu/train.jsonl > transcripts/train_all.jsonl
cat transcripts/valid.jsonl transcripts_ipu/valid.jsonl > transcripts/valid_all.jsonl