Jump Cut Editor (VAD-based)

Automatically remove silences from talking-head videos using neural voice activity detection (Silero VAD). More accurate than FFmpeg silence detection, especially for videos with background noise, breathing sounds, or quiet speech.

Execution Script

execution/jump_cut_vad.py

Quick Start

# Basic silence removal
python3 execution/jump_cut_vad.py input.mp4 output.mp4

# With audio enhancement and color grading
python3 execution/jump_cut_vad.py input.mp4 output.mp4 \
    --enhance-audio \
    --apply-lut .tmp/cinematic.cube

# With "cut cut" restart detection (removes mistakes)
python3 execution/jump_cut_vad.py input.mp4 output.mp4 \
    --detect-restarts \
    --enhance-audio

# Fine-tuned parameters
python3 execution/jump_cut_vad.py input.mp4 output.mp4 \
    --min-silence 0.8 \
    --padding 150 \
    --enhance-audio

What It Does

1. Extracts audio from video as WAV
2. Runs Silero VAD (neural voice activity detection) to identify speech segments
3. Optionally detects "cut cut" restart phrases and removes mistake segments
4. Concatenates speech segments with padding
5. Applies audio enhancement (optional): EQ, compression, loudness normalization
6. Applies color grading (optional): LUT-based color correction

CLI Arguments

input

output

--min-silence

--min-speech

--padding

--merge-gap

--keep-start

--no-keep-start

--enhance-audio

--detect-restarts

--restart-phrase

--whisper-model

--apply-lut

Features

1. Silero VAD (Voice Activity Detection)

Uses a neural network trained specifically for voice detection. Much better than FFmpeg's volume-based silence detection:

2. "Cut Cut" Restart Detection

Say "cut cut" during recording to mark a mistake. The script will:

1. Detect the phrase using Whisper transcription
2. Remove the segment containing "cut cut"
3. Remove the previous segment (where the mistake is)

This lets you redo takes naturally without stopping the recording.

# Enable restart detection
python3 execution/jump_cut_vad.py input.mp4 output.mp4 --detect-restarts

# Custom restart phrase
python3 execution/jump_cut_vad.py input.mp4 output.mp4 \
    --detect-restarts --restart-phrase "start over"

3. Audio Enhancement

Applies a professional voice processing chain:

highpass=f=80            # Remove rumble below 80Hz
lowpass=f=12000          # Remove harsh highs above 12kHz
equalizer (200Hz, -1dB)  # Reduce muddiness
equalizer (3kHz, +2dB)   # Boost presence/clarity
acompressor              # Gentle compression (3:1 ratio)
loudnorm=I=-16           # YouTube loudness standard (-16 LUFS)

python3 execution/jump_cut_vad.py input.mp4 output.mp4 --enhance-audio

4. LUT Color Grading

Apply color grading using standard LUT files:

# Apply .cube LUT
python3 execution/jump_cut_vad.py input.mp4 output.mp4 \
    --apply-lut .tmp/cinematic.cube

Supported formats:

.cube

.3dl

.dat

.m3d

.csp

Parameter Tuning

Silence Detection

--min-silence

--min-silence

--min-speech

--min-speech

Padding

--padding

Recording Workflow

With Restart Detection

1. Start recording
2. Speak naturally
3. Make a mistake → Say "cut cut" → Pause briefly → Redo from checkpoint
4. Continue recording
5. Stop when done

The script automatically removes:

• The segment containing "cut cut"
• The previous segment (your mistake)

Without Restart Detection

1. Start recording
2. Speak with natural pauses
3. Long pauses (>0.5s default) will be cut
4. Finish and run the script

Dependencies

System Requirements

brew install ffmpeg  # macOS

Python Dependencies

pip install torch  # For Silero VAD
pip install whisper  # For restart detection (optional)

Silero VAD is downloaded automatically from torch.hub on first run.

Example Output

🎬 Jump Cut Editor (Silero VAD)
   Input: .tmp/recording.mp4
   Output: .tmp/recording_edited.mp4

📏 Video duration: 180.00s
🎵 Extracting audio...
🎯 Running Silero VAD (min_silence=0.5s, min_speech=0.25s)...
   Found 24 speech segments
     1. 0.00s - 8.32s (8.32s)
     2. 9.12s - 15.44s (6.32s)
     3. 16.28s - 22.60s (6.32s)
     ... and 21 more
📎 After merging close segments: 18 segments
🔲 After adding 100ms padding: 18 segments
📌 Preserving intro: extended first segment to start at 0:00
✂️  Concatenating 18 segments...
🎧 Audio enhancement enabled
✅ Output saved to .tmp/recording_edited.mp4

📊 Stats:
   Original: 180.00s
   New: 142.34s
   Removed: 37.66s (20.9%)

vs. simple_video_edit.py

Use

jump_cut_vad.py

Use

simple_video_edit.py

Troubleshooting

"No speech detected"

• Check that audio track exists in the video
• Try lowering

  --min-speech

  to 0.1

Cuts feel too aggressive

• Increase

  --padding

  (e.g., 150-200)
• Increase

  --min-silence

  (e.g., 0.8)

Breathing sounds being cut

VAD should handle this automatically. If not:

• Increase

  --merge-gap

  to 0.5
• Increase

  --padding

  slightly

Restart detection not finding "cut cut"

• Ensure you speak the phrase clearly
• Try

  --whisper-model medium

  for better accuracy
• Check that Whisper is installed:

  pip install whisper

LUT not applying

• Check file path is correct
• Ensure format is supported (.cube, .3dl, .dat, .m3d, .csp)
• Check FFmpeg has lut3d filter:

  ffmpeg -filters | grep lut3d

Performance

Hardware Encoding (Apple Silicon)

The script automatically uses hardware encoding (

h264_videotoolbox

Benchmark (90-minute 1080p video):

• Software encoding: ~20-25 minutes
• Hardware encoding: ~2-4 minutes

The script checks for hardware encoder availability at startup and caches the result. You'll see either:

• 🚀 Hardware encoding enabled (h264_videotoolbox)

• 💻 Using software encoding (libx264)

Output

• Deliverable: Edited video at specified output path
• Format: MP4 (H.264)
• Encoding: Hardware (10 Mbps) or Software (CRF 18), auto-detected
• Audio: AAC 192kbps (enhanced if

  --enhance-audio

  )
• Resolution/FPS: Matches source