Mortem Trimmer

State-of-the-Art AI-Powered Silence Remover & Noise Reduction Pipeline

Remove silence, kill background noise, and produce broadcast-ready audio/video — all in one command.

Features • Installation • Usage • VAD Engines • Noise Reduction • Examples

---

What is Mortem?

Mortem is a production-grade Python CLI tool that automatically detects and removes silent segments from audio and video files using cutting-edge Voice Activity Detection (VAD) models, then optionally applies deep-learning-based noise reduction to produce crystal-clear output.

Whether you're a YouTuber, podcaster, video editor, or content creator — this tool saves you hours of manual editing by intelligently trimming dead air and cleaning up noisy recordings with a single command.

---

Features

Feature	Description
3 VAD Engines	Silero VAD (neural), Whisper Semantic VAD, Energy-based VAD
2 Noise Reduction Modes	DeepFilterNet3 (SOTA deep learning) & Spectral Gating
Video + Audio Support	Process MP4, MKV, AVI, MOV, WAV, MP3, AAC, and more
CUDA Acceleration	GPU-accelerated inference for blazing-fast processing
Scientific Visualization	Auto-generates speech segment plots (timeline chart)
Speech Manifest	Exports timestamped transcript of all detected speech segments
Zero AV-Drift Splicing	Frame-accurate trim+concat via FFmpeg filter graphs
Strategy Pattern Architecture	Clean, extensible OOP design — plug in your own engines

---

Requirements

• Python 3.10+
• FFmpeg installed and accessible (or provide the path via --ffmpeg)
• CUDA (optional, for GPU acceleration with DeepFilterNet3 and Whisper)

---

Installation

1. Clone the Repository

git clone https://github.com/fadhiilahahmadzikri/autosilenttrimmer.git
cd autosilenttrimmer

2. Create a Virtual Environment

python -m venv .venv

Activate the virtual environment:

• Windows (PowerShell):

  .\.venv\Scripts\Activate.ps1

• Windows (CMD):

  .\.venv\Scripts\activate.bat

• Linux / macOS:

  source .venv/bin/activate

3. Install Dependencies

pip install numpy soundfile torch torchaudio noisereduce tqdm matplotlib seaborn

For DeepFilterNet3 noise reduction (recommended):

pip install deepfilternet

For Whisper-based VAD:

pip install faster-whisper

4. Install FFmpeg

• Windows (Chocolatey):

  choco install ffmpeg

• Linux (apt):

  sudo apt install ffmpeg

• macOS (Homebrew):

  brew install ffmpeg

---

Usage

Basic Syntax

python mortem.py -i <input_file> -o <output_file> [options]

Command-Line Arguments

Argument	Type	Default	Description
-i, --input	str	(required)	Path to the input audio/video file
-o, --output	str	(required)	Path for the output file
-m, --method	str	deepfilter	Noise reduction method: none, spectral, deepfilter
--vad	str	silero	VAD engine: none, energy, whisper, silero
--min-silence	float	0.5	Minimum silence duration (seconds) to trigger a cut
--pad	float	0.2	Padding (seconds) added around each speech boundary
-f, --ffmpeg	str	C:\ProgramData\chocolatey\bin\ffmpeg.exe	Path to FFmpeg executable

---

VAD Engines

Silero VAD (Default — Recommended)

Neural network-based acoustic voice activity detection. Fastest and most reliable for general use.

python mortem.py -i input.mp4 -o output.mp4 --vad silero

Whisper Semantic VAD

Uses OpenAI's Whisper model to transcribe audio and extract word-level timestamps. Best for content where semantic accuracy matters (e.g., podcasts with music).

python mortem.py -i input.mp4 -o output.mp4 --vad whisper

Energy-Based VAD

Lightweight, zero-dependency VAD using signal energy thresholding. Great for clean recordings with obvious silence gaps.

python mortem.py -i input.wav -o output.wav --vad energy

---

Noise Reduction

DeepFilterNet3 (Default — SOTA)

State-of-the-art deep neural network for real-time speech enhancement. Supports CUDA GPU acceleration with chunked processing for O(1) memory usage.

python mortem.py -i input.mp4 -o output.mp4 -m deepfilter

Spectral Gating

Classic signal processing approach using noisereduce library. Lightweight and CPU-friendly.

python mortem.py -i input.wav -o output.wav -m spectral

No Noise Reduction

Skip noise reduction entirely — only perform silence trimming.

python mortem.py -i input.mp4 -o output.mp4 -m none

---

Examples

Remove silence from a video (default settings)

python mortem.py -i raw_video.mp4 -o clean_video.mp4

Podcast cleanup with aggressive silence removal

python mortem.py -i podcast.wav -o podcast_clean.wav --min-silence 0.3 --pad 0.1

Video with only silence trimming (no noise reduction)

python mortem.py -i lecture.mp4 -o lecture_trimmed.mp4 -m none --vad silero

Full pipeline with Whisper VAD + DeepFilterNet3

python mortem.py -i interview.mkv -o interview_clean.mkv -m deepfilter --vad whisper

Audio-only processing with spectral gating

python mortem.py -i recording.wav -o recording_clean.wav -m spectral --vad energy

Custom FFmpeg path

python mortem.py -i input.mp4 -o output.mp4 -f /usr/bin/ffmpeg

---

Output Files

After processing, the tool generates:

File	Description
output.mp4	The final cleaned & trimmed media file
output.mp4.manifest.txt	Timestamped speech segment transcript
output.mp4.plot.png	Scientific visualization of speech vs silence timeline

---

Architecture

The project follows the Strategy Pattern with clean interface abstractions:

INoiseReductionStrategy          IVoiceActivityDetector          IMediaManager
├── DeepFilterNetStrategy        ├── SileroBasedVAD              └── FFmpegMediaManager
├── SpectralGatingStrategy       ├── WhisperBasedVAD
                                 └── EnergyBasedVAD

                    AudioProcessor (Orchestrator)
                    StrategyFactory (Builder)

---

Troubleshooting

Issue	Solution
WinError 127 (DLL conflict)	Run pip uninstall torch torchaudio -y then pip install torch torchaudio
FFmpeg not found	Install FFmpeg or pass the correct path via --ffmpeg
CUDA out of memory	DeepFilterNet3 uses chunked processing — try reducing chunk size or use CPU
Whisper VAD is slow	Use silero VAD instead for faster processing

---

License

This project is open source and available under the MIT License.

---

Star This Repo

If this tool saved you time, please star this repository — it helps others discover it!

---

Built by fadhiilahahmadzikri