Gameplay Clip Extractor

Reduce a 1-hour gameplay capture to a list of 10-second clips worth posting.

When to use

• User has raw .mp4/.mov gameplay footage (their own, no piracy).
• User wants timestamps for highlights without scrubbing manually.
• User wants the actual clipped files written to disk.

Detection signals (combine, don't pick one)

1. Audio peak detection (cheapest, works everywhere)

Gunfire, goal whistles, killstreak voice lines, crowd reactions — all show up as audio amplitude spikes.

ffmpeg -i raw.mp4 -af "silencedetect=noise=-30dB:d=0.5" -f null - 2>&1 \
  | grep silence_end

Use the inverse — long non-silent stretches with sudden RMS jumps are candidates.

Python (cleaner):

import librosa, numpy as np
y, sr = librosa.load("raw.mp4", sr=22050, mono=True)
rms = librosa.feature.rms(y=y, frame_length=2048, hop_length=512)[0]
peaks = np.where(rms > rms.mean() + 2 * rms.std())[0]
peak_times = librosa.frames_to_time(peaks, sr=sr, hop_length=512)

2. Scoreboard / kill-feed OCR (game-specific, higher accuracy)

• BGMI / COD: top-right kill feed → run Tesseract on that ROI every 1s.
• FIFA: scoreboard score-change → diff frames at fixed coords.
• Cache last value; emit event only on change.

3. Pacing heuristic

Cluster nearby peaks (within 3s) into single moments. A real highlight is usually 1 spike, not a noise burst.

Output

[
  {"start": 142.3, "end": 152.8, "label": "kill burst (3 audio peaks)", "score": 0.82},
  {"start": 384.1, "end": 394.6, "label": "score change 1→2", "score": 0.95}
]

Then clip with FFmpeg:

ffmpeg -ss 142.3 -i raw.mp4 -t 10.5 -c copy clips/clip_001.mp4

Heuristics for "is this postable?"

• ≥ 2 audio peaks in a 10s window
• Clip ends on a resolved moment (kill confirmed, goal scored), not mid-action
• No long static frames (loading, respawn screen)

What NOT to do

• Don't auto-publish. Always present clips for human review.
• Don't OCR the entire frame — that's slow and noisy. Define ROIs per game.
• Don't process clips you don't own the rights to.