Video Captions

When to Use

User needs captions or subtitles for video content. Agent handles transcription, timing, formatting, styling, translation, and burn-in across all major formats and platforms.

Quick Reference

Core Rules

1. Engine Selection by Context

Default: Whisper local (turbo model). See engines.md for optional cloud alternatives.

2. Format Selection by Platform

Ask user's target platform if not specified.

3. Professional Timing Standards

Netflix-compliant (default):

• Min duration: 5/6 second (0.833s)
• Max duration: 7 seconds
• Max chars/line: 42
• Max lines: 2
• Gap between subtitles: 2+ frames

Social media:

• Shorter segments (2-4 words)
• More frequent breaks
• Centered or dynamic positioning

4. Segmentation Rules

Break lines:

• After punctuation marks
• Before conjunctions (and, but, or)
• Before prepositions

Never separate:

• Article from noun
• Adjective from noun
• First name from last name
• Verb from subject pronoun
• Auxiliary from verb

5. Word-Level Timestamps

Use word timestamps for:

• Karaoke-style highlighting
• Precise sync verification
• TikTok/Instagram animated captions
• Quality checking transcript accuracy

Enable with --word-timestamps flag.

6. Speaker Identification

For multi-speaker content:

• Use diarization (pyannote local, or cloud APIs if configured)
• Format: [Speaker 1] or [Name] if known
• SDH format: JOHN: What do you think?

7. Quality Verification

Before delivering:

• Check sync at start, middle, end
• Verify character limits per line
• Confirm speaker labels if multi-speaker
• Test burn-in render quality

Workflow

Basic Transcription

# Auto-detect language, output SRT
whisper video.mp4 --model turbo --output_format srt

# Specify language
whisper video.mp4 --model turbo --language es --output_format srt

# Multiple formats
whisper video.mp4 --model turbo --output_format all

Word-Level Timestamps

# Using whisper-timestamped
whisper_timestamped video.mp4 --model large-v3 --output_format srt

# With VAD pre-processing (reduces hallucinations)
whisper_timestamped video.mp4 --vad silero --accurate