lai-youtube

YouTube Workflow

YouTube URL → aligned captions in one command. Requires an API key (/lai-setup) and yt-dlp (pip install yt-dlp).

File path convention

All outputs land in the current directory by default (no subdirectories). Use the YouTube video ID as <base> and chain pipeline files as <base>.<derivation>.<ext>:

<base>.mp4              # downloaded media
<base>.en.vtt           # YouTube auto-caption
<base>.meta.md          # metadata
<base>.aligned.json     # forced-alignment output (when caption.output.format=json)
<base>.srt              # final SRT (when caption.output.format=srt)

<base> is the YouTube ID (e.g. la0CaZ2R8EY); the agent extracts it from the URL.

Basic Command

This command uses nemo_run config syntax (key=value), not -o / --output.

# `<base>` = la0CaZ2R8EY (the YouTube video ID)
lai youtube align "https://youtu.be/la0CaZ2R8EY" \
    caption.output.path=la0CaZ2R8EY.aligned.json \
    caption.output.format=json
# shortcut:
lai-youtube "https://youtu.be/la0CaZ2R8EY" caption.output.path=la0CaZ2R8EY.aligned.json

Add -Y (skip confirmation) for non-interactive runs. caption.output.format is inferred from the path suffix but passing it explicitly is safest. Supported: srt, vtt, ass, json, ttml, lrc, … (see /lai-caption).

Pipeline steps

1. yt-dlp downloads media into the current directory by default (media.output_dir=.); files named <video_id>.mp4 etc.
2. Fetch YouTube captions (or transcribe with Gemini if use_transcription=true)
3. Lattice-1 forced alignment
4. Write in the requested format

Check What's Available First

yt-dlp --list-subs "URL"

• Manual captions → best quality, use defaults
• Auto-captions only → consider caption.input.split_sentence=true if the auto-captions break mid-sentence and you want cleaner segments for downstream karaoke / translation / summarization. Default in examples is false to preserve the source's original cue boundaries
• No captions → add use_transcription=true (needs a Gemini key — see /lai-transcribe)

Common Options

All use key=value syntax, dot-nested:

• use_transcription=true — transcribe with Gemini instead of using YouTube captions
• caption.input.split_sentence=false (default in examples) — preserve source cue boundaries. Set to true to re-segment into clean sentences (helpful for messy auto-captions)
• caption.render.word_level=true — per-word timestamps (needed for karaoke; pair with a JSON or ASS output)
• caption.output.path=<base>.aligned.json / caption.output.format=json — output destination (use <base>.<derivation>.<ext> naming)
• media.prefer_audio=true — audio-only download (faster, smaller)
• media.prefer_audio=false — keep the video. Pair with media.output_format=mp4 explicitly; prefer_audio=false alone may still emit audio-only output, which breaks downstream karaoke (no width/height for probe_media.py)
• media.output_dir=. — where downloaded audio / vtt land (default: current directory)
• media.streaming_chunk_secs=300 — long-video chunking

Full list: lai youtube align --help. Discover the full config tree with lai youtube align --to-yaml /tmp/cfg.yaml URL (exports and exits without running).

Download Only (lai youtube download)

When you need just the media (no alignment), or want to manage the media + caption files yourself:

lai youtube download --direct -Y \
    yt_url="https://youtu.be/la0CaZ2R8EY" \
    media.output_dir=. \
    media.output_format=mp4 \
    media.prefer_audio=false

• media.output_format: mp4 (default with prefer_audio=false), mp3, wav, …
• only=meta — re-download metadata only (writes <id>.meta.md, no media)

The downloaded files follow the pattern <id>.mp4, <id>.en.vtt, <id>.meta.md inside media.output_dir (current directory by default). Pipe the pair into /lai-align afterwards.

Karaoke Recipe (end-to-end)

<base> = YouTube video ID (e.g. la0CaZ2R8EY):

# Step 1 — YouTube → aligned JSON with word-level timestamps
lai youtube align -Y "https://youtu.be/la0CaZ2R8EY" \
    caption.output.path=la0CaZ2R8EY.aligned.json \
    caption.output.format=json \
    caption.input.split_sentence=false \
    caption.render.word_level=true

# Step 2 — JSON → ASS with per-word karaoke highlighting (see /lai-caption)
laicap-convert -Y la0CaZ2R8EY.aligned.json la0CaZ2R8EY.karaoke.ass \
    ass.karaoke_effect=sweep \
    ass.karaoke_color_scheme=azure-gold

(render.word_level is not needed on the convert side — it's default word-scope. The upstream caption.render.word_level=true in Step 1 is what populates the words arrays that karaoke reads.)

For bilingual karaoke, run /lai-translate on la0CaZ2R8EY.aligned.json first (it preserves words), then run step 2 on the bilingual JSON.

Operational Tips

For long alignments (Step 3 lattice search can take minutes on long-form video), redirect to a log file so you can see EXIT code and stderr after truncation:

lai youtube align -Y URL caption.output.path=la0CaZ2R8EY.aligned.json ... > la0CaZ2R8EY.align.log 2>&1
echo "EXIT=$?"
tail -50 la0CaZ2R8EY.align.log

Common Issues

Problem	Fix
yt-dlp not installed	pip install yt-dlp
Video private / unavailable	URL must be publicly accessible
No captions and no Gemini key	Set one up — see /lai-transcribe
Region-restricted	Use a VPN or fetch from an unrestricted region
Age-restricted	yt-dlp --cookies-from-browser chrome to download, then /lai-align
API key invalid	/lai-setup
media.prefer_audio=false but downloaded file is still mp3 / m4a	Add media.output_format=mp4 explicitly — they're separate flags
Couldn't load entrypoint youtube: No module named 'lattifai' on lai startup	Stale editable install — see /lai-setup Common Issues

Related Skills

• /lai-setup — first-time install and auth
• /lai-align — when media + captions are already local
• /lai-transcribe — transcribe without alignment (Gemini accepts YouTube URLs too)
• /lai-diarize — speaker labels on aligned output
• /lai-translate — translate captions
• /lai-caption — convert output format