Youtube Bilingual Subtitle Delivery

YouTube Bilingual Subtitle Delivery

Use when the user gives a YouTube URL and wants an MP4 with English/Chinese subtitles.

When this skill is especially useful

• YouTube downloads as webm and must become mp4
• English auto-captions are available but Chinese captions fail or get HTTP 429
• You need a practical delivery result more than a perfect fully-automated translation pipeline
• You want both .srt, softsub .mp4, and hardcode .mp4

Output standard

Deliver in this order: 1. bilingual .srt 2. softsub .mp4 3. hardcode .mp4

Use self-descriptive names under ~/.Hermes/workspace/output/<task-folder>/.

Proven workflow

1) Prepare output folder

mkdir -p ~/.Hermes/workspace/output/<topic-folder>

2) Download the YouTube source

Prefer yt-dlp with a descriptive filename.

yt-dlp --no-playlist -o '%(title).120s [%(id)s].%(ext)s' '<youtube-url>'

3) Convert to MP4 if needed

If the merged output is webm, transcode to H.264/AAC MP4.

ffmpeg -y -i input.webm -c:v libx264 -preset fast -crf 18 -c:a aac -b:a 192k -movflags +faststart output.mp4

Verify the cut duration before subtitle work:

ffprobe -v error -show_entries format=duration:stream=codec_name,codec_type,width,height -of json output.mp4

4) Check subtitle availability first

List available YouTube subtitle tracks:

yt-dlp --list-subs '<youtube-url>'

If English auto-captions exist, try downloading them first:

yt-dlp --skip-download --write-auto-sub --sub-langs 'en-orig,en' --sub-format srt -o '<basename>.%(ext)s' '<youtube-url>'

If Chinese auto-captions also exist, try:

yt-dlp --skip-download --write-auto-sub --sub-langs 'zh-Hans,zh-Hant' --sub-format srt -o '<basename>.%(ext)s' '<youtube-url>'

Important fallback: Chinese caption download may fail with HTTP 429

This happened in practice even when --list-subs showed Chinese tracks. If that happens:

• keep the English auto-caption time axis
• create Chinese lines manually and attach them to the same subtitle event
• do not block the delivery waiting on the Chinese download endpoint to recover

Direct JSON3 caption URL fallback from metadata

If yt-dlp --write-auto-subs fails with HTTP 429 but yt-dlp --dump-single-json succeeded, do not immediately fall back to Whisper. The metadata often contains usable automatic_captions[lang] track URLs. Fetch the json3 URL directly for the primary language and parse it into a timing scaffold.

import json, urllib.request, pathlib
m = json.load(open('video_metadata.json'))
for lang in ['en-orig', 'en', 'zh-Hans', 'zh-Hant']:
    tracks = (m.get('automatic_captions') or {}).get(lang) or []
    js = next((t for t in tracks if t.get('ext') == 'json3'), None)
    if not js:
        continue
    req = urllib.request.Request(js['url'], headers={'User-Agent': 'Mozilla/5.0'})
    try:
        data = urllib.request.urlopen(req, timeout=30).read()
        pathlib.Path(f'captions_{lang}.json3').write_bytes(data)
        print(lang, len(data))
    except Exception as e:
        print('caption fetch failed', lang, e)

Parse JSON3 events[*].tStartMs, dDurationMs, and segs[*].utf8 into SRT events, then group tiny auto-caption events into readable chunks before translation. This is especially useful when English JSON3 downloads work but translated Chinese tracks still return 429.

Important fallback: some YouTube videos have no usable captions at all

This happened with an official music video where yt-dlp --list-subs showed no automatic captions, only live_chat.

If there is no usable English subtitle track: 1. extract mono 16 kHz WAV from the MP4 2. run local whisper to create an English timing scaffold 3. rewrite the raw ASR into grouped lyric/phrase blocks 4. attach Chinese manually on the same event timeline

Example:

ffmpeg -y -i source.mp4 -vn -ac 1 -ar 16000 audio.wav
whisper audio.wav --model turbo --language en --task transcribe --output_format srt --output_dir whisper_out

Guideline:

• use Whisper for timing, not for final lyric wording
• for songs, aggressively regroup tiny ASR fragments into readable musical lines
• if the ending includes applause / spoken thanks / trailing ambience, trim or rewrite those blocks before calling the SRT final

Important fallback: external Argos/translation pipeline may fail

A practical failure encountered:

• a skill-local Argos setup tried to install dependencies that pulled spacy
• the environment landed on Python 3.9 and failed resolving thinc / spacy
• result: the fully-local translation path was not reliable enough for immediate delivery

Guideline:

• if automated translation setup becomes dependency-heavy or version-fragile, do not stall the task
• use the English timing baseline plus manual Chinese to finish the deliverable
• only invest in the local translation stack when the user specifically wants an automated reusable pipeline

5) Build a clean bilingual SRT

If the English auto-captions are messy, treat them as a timing scaffold and rewrite the text into grouped lyric/phrase blocks.

Keep format per subtitle event:

• English on top line
• Chinese on bottom line

Example block:

1
00:00:19,439 --> 00:00:31,960
A thousand generations / Falling down in worship / To sing the song of ages to the Lamb
千代万代 / 俯伏敬拜 / 向羔羊唱那亘古之歌

6) Audit bilingual completeness

Before calling anything final, confirm there are no English-only subtitle blocks.

If you have the audit script from the video-bilingual-subtitle-delivery skill:

python scripts/audit_bilingual_srt.py /path/final_bilingual.srt

Required result:

• english_only=0
• chinese_only=0
• empty=0

7) Export softsub MP4 first

ffmpeg -y -i source.mp4 -i final_bilingual.srt \
  -map 0:v -map 0:a -map 1:0 \
  -c:v copy -c:a copy -c:s mov_text \
  -metadata:s:s:0 language=eng \
  output_softsub.mp4

Verify streams:

ffprobe -v error -show_entries format=duration:stream=codec_type,codec_name:stream_tags=language -of json output_softsub.mp4

8) Export hardcode MP4

If you have the Python hardcode renderer from the video-bilingual-subtitle-delivery workflow, use it rather than relying on ffmpeg subtitle filters.

Example:

python hardcode_bilingual_srt.py --video source.mp4 --srt final_bilingual.srt --output output_hardcode.mp4

Important pitfall: hardcode renderer may extend duration

Observed in practice:

• source MP4 duration: 307.444s
• hardcoded output duration: 312.667s

Cause: overlay concat generation can leave an oversized tail.

Fix: trim the hardcoded result back to the source duration.

ffmpeg -y -i output_hardcode.mp4 -t <source_duration_seconds> -c copy -movflags +faststart output_hardcode_trimmed.mp4

Then verify source vs final durations:

ffprobe -v error -show_entries format=duration -of default=nw=1:nk=1 source.mp4
ffprobe -v error -show_entries format=duration -of default=nw=1:nk=1 output_hardcode_trimmed.mp4

A tiny container rounding difference is acceptable; multi-second drift is not.

9) Visual verification

Extract 1–2 frames from moments with visible subtitles and inspect them.

ffmpeg -y -ss 00:00:25 -i output_hardcode_trimmed.mp4 -frames:v 1 check1.png
ffmpeg -y -ss 00:04:30 -i output_hardcode_trimmed.mp4 -frames:v 1 check2.png

Confirm:

• bilingual subtitles are visibly burned in
• English is top line
• Chinese is bottom line
• no missing Chinese on checked frames

Important pitfall: YouTube auto-caption timing is often unreliable

In practice, YouTube's auto-generated caption timestamps are frequently misaligned — words appear too early or too late relative to speech. When the user reports that subtitles don't sync with the audio ("字幕和时间轴对不上"), the root cause is almost always the YouTube auto-caption timing, not the hardcode renderer.

Fix: Replace YouTube auto-captions with local Whisper transcription for word-level timing accuracy:

# Extract audio
ffmpeg -y -i source.mp4 -vn -ac 1 -ar 16000 audio.wav

# Transcribe with Whisper (turbo model, ~30s for a 5-min video)
whisper audio.wav --model turbo --language en --task transcribe --output_format srt --output_dir whisper_out/

Then parse the Whisper SRT, group into readable subtitle chunks (~12 words or sentence-ending), translate with deep_translator, and build the bilingual SRT.

Practical guidance

• YouTube auto-captions are fine for quick transcript work, but NOT for timing-critical subtitle delivery. When precision matters, always rebuild with local Whisper.
• For music videos, audit caption text quality even when --list-subs shows en-orig or other auto tracks. Auto/ASR captions can contain badly mangled lyrics (for example repeated hooks rewritten as unrelated words). Treat them as disposable if lyrics are wrong.
• For lyric videos, use Whisper primarily as a timing scaffold, then manually correct the English against known/repeated lyrics before translating. Preserve repeated hooks consistently and regroup lines musically/readably instead of preserving tiny raw ASR fragments.
• If the user needs a polished worship/song translation, manually refine Chinese for natural phrasing rather than shipping raw machine translation.
• Don't let toolchain fragility block delivery; finish the user-facing asset first.

Deliverable checklist

Do not call the task done until all are true:

• source is confirmed and in MP4 if requested
• bilingual SRT exists
• bilingual audit passes with no English-only blocks
• softsub MP4 exists and probes cleanly
• hardcode MP4 exists and duration matches source closely
• at least one visual spot-check confirms subtitles are actually burned in