evo-video-silence-remover

Video Silence Remover Skill

Overview

Removes non-teaching content (openings, pauses) from teaching videos by combining audio energy analysis with visual frame differencing. All decision thresholds are derived from the input data distribution at runtime.

Scripts

• scripts/utils.py – audio/frame extraction, RMS energy, frame diffs, duration queries
• scripts/detect_segments.py – opening and pause detection with data-driven thresholds
• scripts/edit_video.py – ffmpeg segment extraction, concatenation, report generation
• scripts/run_pipeline.py – end-to-end orchestration and output validation

Usage

import sys
sys.path.insert(0, '/app/environment/skills/evo-video-silence-remover/scripts')
from run_pipeline import run_pipeline, validate_output

report = run_pipeline(
    input_video='data/input_video.mp4',
    output_video='compressed_video.mp4',
    report_path='compression_report.json',
    window_sec=0.25,
    min_pause_sec=2.0,
)

ok, issues = validate_output('compressed_video.mp4', 'compression_report.json')
if ok:
    print("All checks passed!")
else:
    for issue in issues:
        print(f"Issue: {issue}")

How It Works

Opening Detection

1. Extracts grayscale frames and computes mean absolute differences per block.
2. Derives a static-vs-active threshold from the frame-diff distribution using a gap-analysis method on the sorted lower half of block means.
3. Identifies the contiguous truly-static prefix from the video start.
4. Searches for the largest relative audio energy drop near the static boundary to locate the transition into teaching content.
5. Refines the boundary to the nearest local energy minimum.

Pause Detection

Three complementary strategies scan the teaching portion of the audio:

• Adaptive local windows – compares energy to a threshold derived from local and global statistics so that quiet teaching regions do not mask pauses.
• Global percentile – uses the midpoint of the 10th and 25th percentiles of the teaching-content energy distribution.
• Running-median ratio – flags sustained dips below a data-derived fraction of the smoothed energy baseline.