Video Silence Trimmer

Video Silence Trimmer is a small desktop-style tool for cleaning up long recordings by cutting dead air out of videos in batches.

It is built for the kind of files people actually collect over time: recorded lectures, class sessions, screen recordings, training videos, meeting recordings, interview captures, and rough content that has too much waiting, pausing, or empty space.

The Problem

A lot of useful recordings are much longer than they need to be.

Examples:

• a classroom lecture where the teacher pauses for long stretches while writing on the board
• a screen recording with long setup delays or waiting time
• a meeting or interview recording with repeated dead air
• study material recorded on a phone with uneven pacing

These files are still useful, but they are slower to review, harder to share, and frustrating to revisit.

Why Not Just Use FFmpeg Directly

• ffmpeg's silencedetect filter works on amplitude threshold alone. It has no concept of speech — it will cut a quiet voice and keep a loud background hum. For spoken recordings, this produces unusable output without manual tuning per file.

• silero-vad solves this by detecting speech probability per frame, not audio energy. The tradeoff is runtime — it is slower and requires a model download on first use.

• This tool wraps both, lets the user choose, and handles the stitching, batching, and output management that neither provides.

• This tool will/can be improved to functionalise auto content tagging to video sections to allow bookmarking segments with topics it is related to.

The Solution

This tool scans a folder of videos, detects silence or non-speech sections, and creates shorter trimmed versions in a _trimmed_output folder.

It offers two detection modes:

• ffmpeg: a simpler audio-threshold based detector
• silero-vad: a speech-aware detector that is usually better for spoken recordings

It also has a Textual TUI so the user can:

• pick a folder
• choose the backend
• tune silence settings
• watch progress live
• review results after processing

Common Use Cases

• recorded lectures from class
• tutoring sessions
• online course screen recordings
• meeting recordings
• interview footage
• self-recorded study explanations
• raw video notes that need cleanup before sharing

Requirements

The intended requirement is just:

• Python 3.10+

The Windows launcher then takes care of the rest automatically.

How To Run

From the project root on Windows:

.\silence_trimmer.bat

That opens the TUI after setup is complete.

If you want to run the CLI directly:

.\.venv_trimmer\Scripts\python.exe -m silence_trimmer --cli "<path\to\videos>"

What The Build Script Installs Automatically

The launcher in silence_trimmer.bat is designed so the user does not have to manually install project dependencies one by one.

On first run it will:

• create or repair the local virtual environment at .venv_trimmer
• install the core Python packages for the app and TUI
• provision local ffmpeg and ffprobe into tools/ffmpeg if they are not already available
• install the Silero runtime dependencies
• download and keep a local silero-vad copy beside the repo
• install the tagging dependencies used for transcription and topic extraction

On later runs it reuses what is already present instead of downloading everything again.

Output

Processed files are written into a sibling output folder inside the chosen input directory:

input_folder/
├── <output_01>.mp4
├── <output_02>.mkv
└── _trimmed_output/
    ├── output_01_trimmed.mp4
    ├── output_02_trimmed.mp4
    └── _session_manifest.json

The trimmed file is still a normal video with audio for the kept sections. The silent parts are removed; the remaining video and audio are stitched together.

Architecture

Input Folder
    │
    ▼
Folder Scanner → filters non-video, skips video-only files
    │
    ▼
Detection Backend (user choice)
    ├── ffmpeg silencedetect (fast, threshold-based)
    └── silero-vad (speech-aware, model-based)
    │
    ▼
Segment Builder → computes keep/cut intervals
    │
    ▼
ffmpeg Concat → stitches kept segments into output file
    │
    ▼
_trimmed_output/
    ├── trimmed video files
    └── _session_manifest.json (per-file segment log)

TUI layer (Textual) runs over this pipeline and manages folder selection, backend config, progress display, and results review.

Screenshots

Main TUI

Processing / Results

Notes

• silero-vad works best for spoken content.
• Video-only files cannot be analyzed for silence and are skipped.
• If tagging is enabled, the transcription model may still download its own model files on first use for the selected model size.

Limitations

• Windows only currently (bat launcher + venv path assumptions)
• silero-vad requires internet on first run to download model weights
• Video-only files (no audio track) are skipped entirely
• Very short silence gaps below ~300ms are not removed to avoid cutting mid-word pauses
• LLM topic tagging requires transcription first — adds significant runtime for large files
• No GPU acceleration — silero runs on CPU only in this build

License

This project is licensed under the MIT License. See LICENSE.