video-bilingual-subtitle-delivery

Video Bilingual Subtitle Delivery

Produce bilingual subtitle deliverables in a strict order: get the right source cut, lock English timing to speech, attach Chinese to the same subtitle event, audit for missing Chinese, then export softsub and hardcode outputs.

Quick start

1. Verify the source clip duration with ffprobe.
2. Confirm whether the current subtitle file belongs to this exact cut.
3. If timing is bad, rebuild English timing first; do not guess with large global offsets.
4. After timing is good, attach or refine Chinese on the same time axis.
5. Audit the SRT for English-only blocks with scripts/audit_bilingual_srt.py.
6. Export softsub first. Export hardcode last.

Local Whisper pipeline

Use the bundled builder when you want a no-OpenAI baseline pipeline:

python scripts/build_bilingual_subtitles.py \
  --video /path/input.mp4 \
  --output-dir /path/output_dir \
  --basename topic_name \
  --whisper-model turbo \
  --translate-backend argos

What this does:

• extracts mono 16 kHz audio from the source video
• runs local whisper CLI to create an English timing baseline
• groups short ASR lines into more readable subtitle events
• translates with a pluggable backend
• applies lightweight Chinese polishing by default for machine-translated output
• writes a bilingual SRT on the same time axis

Current translation backends:

• manual — no API key required; writes 【待补中文】... placeholders under each English line so timing/editing can continue locally
• none — English-only output on the grouped time axis
• argos — fully local offline translation using Argos Translate (en -> zh)

⚠️ Translation backend reliability: prefer raw Google Translate API over deep_translator

Problem: deep_translator's translate_batch() consistently gets rate-limited on large SRTs (800+ texts), timing out or returning empty results after a few batches. Individual translate() calls are also slow (~2-3s per text). The googletrans library may not be installable in the active Python environment.