youtube-clip-transcribe

name	youtube-clip-transcribe
description	Downloads YouTube sermon audio, crops by timestamps, optimises volume (quiet or hot), merges multi-part clips, transcribes with praisonai-audio-editor, maps to YAML decks, and packs deliverables. Use when the user gives a YouTube URL with start/end times, full video, two-part merge, sermon transcript, or BIC sermon deck pack.

YouTube clip → normalise → transcribe → pack

Run every step on the user's Mac. Do not stop at instructions.

Repeat the pipeline per clip. Never merge transcripts unless the user explicitly asks to join two videos into one audio file first.

Prerequisites

Tool	Check
ffmpeg / ffprobe	ffmpeg -version
OPENAI_API_KEY	bash -lc 'test -n "$OPENAI_API_KEY" && echo ok'
This repo	~/praisonai-audio-editor
yt-dlp	conda python3 -m yt_dlp (homebrew yt-dlp often returns 403)
YAML decks	/Users/praison/praisonaippt/examples/*.yaml

Shell: use bash -lc '…' so ~/.bashrc loads OPENAI_API_KEY.

Parse user input

1. YouTube URL → {VIDEO_ID}
2. Start / End — -to is absolute on the full timeline
3. "till the end" — -ss START only (no -to)
4. Full video — skip crop; use ~/Downloads/{VIDEO_ID}_full.m4a
5. Two-part sermon — crop each, merge, then normalise → transcribe once

Output naming

Artifact	Path
Full download	~/Downloads/{VIDEO_ID}_full.m4a
Cropped / normalised	~/praisonai-audio-editor/{stem}.m4a
Transcript JSON	…/{stem}.transcript.json
Plain text	…/{stem}.transcript.txt
Log	…/{stem}_transcribe.log

Checklist

- [ ] Download (conda yt-dlp)
- [ ] Crop or merge parts
- [ ] Normalise volume — report mean/max before & after
- [ ] ffprobe duration check
- [ ] transcribe --format json
- [ ] extract-text
- [ ] Pack + YAML map (if requested)

Step 1 — Download

zsh -c "source $(conda info --base)/etc/profile.d/conda.sh && conda activate test && \
  python3 -m yt_dlp -f 'ba/b' --no-playlist --concurrent-fragments 8 --no-part \
  -x --audio-format m4a \
  -o '/Users/praison/Downloads/{VIDEO_ID}_full.%(ext)s' 'URL'"

403 → pip install -U yt-dlp + --extractor-args 'youtube:player_client=android,web'.
No --print on download — it can skip writing the file.

Step 2 — Crop

ffmpeg -y -nostdin -i FULL.m4a -ss START -to END -map 0:a:0 -c copy OUT.m4a
# end of video: omit -to

Step 2b — Volume optimisation

Target: −16 LUFS, −1.5 dBTP.

Condition	Action
Too quiet (mean < −22 dB or max < −8 dB)	Auto loudnorm
Optimal (−22…−14 mean, −8…−2 max)	Skip — "Volume OK"
Too loud (mean > −12 dB or max ≥ −1 dB)	--in-place --force

cd ~/praisonai-audio-editor
python3 -m praisonai_editor normalize "$AUDIO" --in-place
python3 -m praisonai_editor normalize "$AUDIO" --in-place --force   # hot audio

Wrapper: scripts/optimize-audio-volume.sh

Step 2c — Merge two parts

ffmpeg -y -i PART1.m4a -i PART2.m4a \
  -filter_complex "[0:a][1:a]concat=n=2:v=0:a=1[out]" -map "[out]" \
  -c:a aac -b:a 192k MERGED.m4a

Normalise merged file, then transcribe once.

Step 3 — Transcribe

Full local pipeline (silence cut + normalise + transcribe): sermon-transcribe — run all steps when user provides audio; this YouTube skill covers download + crop + normalise only.

Transcribe rules (Tamil+English, --speed 2, sanity checks): sermon-transcribe Steps 6–10.

Mixed Tamil/English: omit --language. English-dominant: --language en.

bash -lc 'cd ~/praisonai-audio-editor && python3 -m praisonai_editor transcribe \
  FILE.m4a --format json \
  2>&1 | tee STEM_transcribe.log'

Optional: --speed 2 after 5 min quality test (see sermon-transcribe).

Step 4 — Extract text

python3 -m praisonai_editor extract-text STEM.transcript.json

→ {stem}.transcript.txt

Step 5–6 — YAML map + deck pack

Map stem → examples/*.yaml by topic/verses. Pack script:

python3 /Users/praison/praisonaippt/scripts/build_sermon_pack2.py

Output: ~/Downloads/BIC-Sermon-Deck-Pack-2/ + sermon_video_map.json

Do not use

• Homebrew yt-dlp alone (403)
• --download-sections on long clips
• transcribe --force-transcribe

More detail

Canonical skill: ~/.cursor/skills/youtube-clip-transcribe/SKILL.md (reference.md, examples.md)