denoise-video

Overview

denoise-video is a safety-first CLI tool and Agent Skill for removing background noise from finished videos (MP4, MOV, MKV) without sacrificing picture quality.

Traditional video editing tools or online converters often re-encode the entire video stream during noise reduction—causing quality degradation, color space drift, and slow processing times. denoise-video separates video and audio tracks, enhances the speech audio using neural models strictly on your local machine, and remuxes the original video stream (c:v copy) untouched.

Key Highlights

• ⚡ Zero Picture Re-encoding: Copies the original video stream (c:v copy). Video resolution, bitrate, color space, and frame timing remain 100% bit-exact.
• 🔒 100% Local & Private: No cloud uploads, external API dependencies, or telemetry. Media files never leave your system.
• 🧠 Neural Speech Enhancement: Leverages DeepFilterNet3 to suppress complex background noise (street traffic, fans, air conditioning, room reverb) while keeping speech natural.
• 🛡️ SHA-256 Stream Integrity Verification: Hashes video frame streams before and after remuxing to guarantee bit-for-bit picture preservation before creating the output file.
• 🧩 Pluggable Architecture: Supports DeepFilterNet3, custom local CLI audio models (e.g., custom PyTorch/ONNX models), and lightweight FFmpeg afftdn fallbacks.
• ⚠️ Non-Destructive Guarantee: Refuses to overwrite source files or pre-existing output destinations.

Architecture & How It Works

1. Probe & Sanity Check: Validates input format with ffprobe. Enforces exactly 1 video stream and 1 mono/stereo audio stream, with start times synchronized within 50 ms.
2. Track Demux & Parallel Execution:
   • Video Track: Directly passed through using ffmpeg -c:v copy.
   • Audio Track: Extracted to 48 kHz PCM WAV, processed by the selected denoise engine, and re-encoded to AAC.
3. Remux & SHA-256 Verification: Combines the original video stream and denoised audio. Hashes the output video stream to verify 100% identity with the source before placing the final file.

Denoise Backend Comparison

Prerequisites

• Python: 3.9+
• FFmpeg & FFprobe: Must be installed and available in system PATH.

  # macOS
  brew install ffmpeg
  
  # Ubuntu/Debian
  sudo apt install ffmpeg
  

Install as an Agent Skill

Clone the repository and symlink its skill directory into your agent's skills folder, using the path your runtime expects:

git clone https://github.com/manwithshit/denoise-video.git

# Claude Code
ln -s "$(pwd)/denoise-video/skill" ~/.claude/skills/denoise-video

# Codex / other agent runtimes
ln -s "$(pwd)/denoise-video/skill" ~/.agents/skills/denoise-video

The symlink's name (denoise-video) becomes the installed skill directory name, matching the name field in SKILL.md.

Note: skill/agents/openai.yaml is an optional interface manifest (display name, one-line description, default prompt) for OpenAI-compatible agent runtimes. Claude Code reads SKILL.md directly and doesn't need it.

Usage

1. Check Available Backends

Inspect available tools (ffmpeg, ffprobe, uv, deepFilter):

python3 skill/scripts/denoise_video.py check

2. Run Denoising (DeepFilterNet3)

Run noise reduction using DeepFilterNet3 with standard attenuation (18 dB cap):

python3 skill/scripts/denoise_video.py run "/path/to/video.mp4" \
  --backend deepfilternet3 \
  --strength standard \
  --allow-download

Attenuation Strength Presets

• natural: Caps attenuation at 12 dB (preserves background ambience).
• standard (Default): Caps attenuation at 18 dB (balanced clarity and natural speech).
• strong: Unrestricted model attenuation (maximum noise removal).

Note: Pass --allow-download to grant consent for downloading the DeepFilterNet3 checkpoint on first run.

3. Run with a Custom Local Model

Use an existing local CLI model that takes PCM WAV input and writes PCM WAV output:

python3 skill/scripts/denoise_video.py run "/path/to/video.mp4" \
  --backend custom \
  --custom-command 'my-enhancer --input "{input}" --output "{output}"'

4. Run FFmpeg Fallback

Use the lightweight, offline FFmpeg afftdn filter:

python3 skill/scripts/denoise_video.py run "/path/to/video.mp4" --backend ffmpeg

Custom Output Destination

By default, output is saved to <name>_denoised.<ext> in the same directory. Specify a custom path with --output:

python3 skill/scripts/denoise_video.py run "/path/to/video.mp4" --output "/path/to/output.mp4"

Safety & Boundaries

• Track Constraints: Rejects files with multiple video/audio streams, subtitles, data streams, or attachments to prevent accidental track dropping.
• Audio Constraints: Supports mono or stereo audio streams. Multi-channel surround audio (5.1/7.1) is refused.
• Sync Guard: Enforces audio and video start time alignment within 50 ms.
• No Overwrites: Refuses to overwrite source files or existing output paths.

Credits & License

The neural workflow uses DeepFilterNet by Hendrik Schröter (MIT/Apache-2.0). Dual-licensed under MIT / Apache-2.0.