create-jump-cuts

create-jump-cuts

Purpose

Remove inter-word and inter-sentence pauses from a talking-head clip so the delivery is tight and creator-native. Gaps above a configurable threshold are deleted; a small pad of audio/video is kept around each speech segment to preserve natural sentence starts and endings. The result is a re-encoded single video with all keep-segments concatenated.

This is the jump-cut transition type described in CREATOR_GRAMMAR.md §3 — the standard breath-removal edit for the creator-talking-head archetype.

Inputs

• --source (required) — path to source video file.
• --words (mutually exclusive with --auto) — path to a word-timestamps JSON file in the shape output by transcribe-audio-fal:

  [{"text": "Hey", "start": 0.12, "end": 0.45}, ...]
  

• --auto (mutually exclusive with --words) — derive speech segments automatically using ffmpeg silencedetect (no transcript needed).
• --gap-threshold (optional, default 0.35) — gaps longer than this many seconds between speech are candidates for removal.
• --pad (optional, default 0.08) — seconds of audio/video to keep on each side of a speech segment (prevents clipping first/last phoneme).
• --output (required) — path for the output video file (.mp4).
• --crf (optional, default 18) — x264 quality.

Workflow

1. Verify ffmpeg and ffprobe are on PATH.
2. Build speech segments:
   • --words mode: merge consecutive word timestamps where inter-word gap ≤ gap-threshold into contiguous speech blocks. Apply pad to each block's start/end (clamped to clip bounds).
   • --auto mode: run ffmpeg -af silencedetect=noise=-35dB:d=<gap-threshold> on the source, parse silence_start / silence_end lines from stderr, invert to get speech segments. Apply pad.
3. If no speech segments are found, copy source to output unchanged and write manifest with a warning.
4. For each keep-segment: trim with ffmpeg -ss <start> -i <source> -t <dur> -c:v libx264 -crf <crf> -preset fast -pix_fmt yuv420p -c:a aac -b:a 192k.
5. Write an ffmpeg concat list (concat.txt) and run ffmpeg -f concat -safe 0 -i concat.txt -c copy to join.
6. Clean up temporary segment files.
7. Write manifest.json to the same directory as --output listing every removed range and total time removed.

Run:

python3 skills/atoms/editing/create-jump-cuts/scripts/jump_cut.py \
  --source /path/to/raw.mp4 \
  (--words /path/to/words.json | --auto) \
  --output /path/to/tight.mp4 \
  [--gap-threshold 0.35] [--pad 0.08] [--crf 18]

Output

• <output>.mp4 — re-encoded video with pauses removed.
• manifest.json (same directory as output) — includes:
  • skill_name, run_id, input_path, output_files, provider, model_or_tool, created_at, duration_seconds, status, warnings, errors
  • removed_ranges — array of {start, end, duration_s} for every deleted gap
  • total_removed_seconds — sum of all removed gap durations
  • source_duration_seconds — original clip length
  • output_duration_seconds — final clip length

Quality Checks

• Output file exists and is non-zero size.
• ffprobe confirms output is playable H.264/yuv420p.
• Output duration ≈ source duration minus total_removed_seconds (±0.3s tolerance for pad/encode rounding).
• manifest.json lists all removed ranges and the total savings.
• No speech content is cut — only silence/pause gaps above threshold.
• If --words is provided, every word boundary is preserved within pad tolerance.

Failure Modes

• ffmpeg / ffprobe not installed — install via brew install ffmpeg.
• Source missing or unreadable — script exits with clear error before processing.
• No silence detected (--auto mode on a clip with no pauses) — output equals source; manifest records warning, status pass.
• Mutually exclusive flags both provided (--words + --auto) — script exits with argument error.
• Corrupt or empty words JSON — script exits with clear parse error.
• Concat step fails — individual segment files are left in place for debugging; manifest records error.
• Very short source (<1s) — processed normally; output may equal source if no gaps found.