Tired of manually cutting out silent pauses in your videos? VideoSpeeder is here to revolutionize your workflow! ⚡️

This tool intelligently analyzes your video, detects silent segments, and automatically speeds them up, saving you precious editing time. Perfect for:

• 🎙️ Podcasters & YouTubers: Condense long recordings quickly.
• 👨‍🏫 Educators & Presenters: Make lectures more engaging.
• 💻 Coders & Tutorial Makers: Skip the thinking pauses in screen recordings.
• ...anyone who wants faster, tighter videos!

🎥 Have a look at a video that uses it here.

• 🗣️ Speech Detection (VAD): Uses Silero VAD (on by default) to detect actual human speech — ignores wind, keyboard, mouse, and ambient noise.
• 🤫 Fallback Silence Detection: --no-vad mode uses ffmpeg's silencedetect filter for amplitude-based detection (useful for clean audio).
• ⏩ Dynamic Speed-Up: Automatically calculates the optimal speed for silent parts, aiming for a concise ~4-second duration while capping at a blazing 1000x! Shorter silences get a fixed 4x boost.
• 📊 Rich CLI Stats: Get beautifully formatted video stats upfront using rich.
• 👁️ Visual Speed Indicator: Optional >> [Speed]x overlay shows exactly when the video is sped up.
• 💡 Transcription Power: Includes a separate script (transcribe.py) leveraging OpenAI Whisper for highly accurate VTT subtitle generation. Choose your model size!
• ⚙️ Fine-Tuned Control: Adjust silence detection threshold (dB) and minimum duration (seconds).
• ⏱️ Segment Processing: Process only specific parts of your video using --offset and --process-duration.
• 🚀 GPU Acceleration: Supports NVIDIA NVENC for encoding and CUVID/NVDEC for decoding (requires compatible hardware, drivers, and FFmpeg build) for significantly faster processing.
• ⏳ Progress Bar: Keep track of the process with a tqdm progress bar.

Just is a command runner. Install it, then:

# 1. Set up the Python venv (installs torch, silero-vad, etc.)
just venv

# 2. Speed up a single video (VAD + GPU on by default)
just speed input=my_video.mp4 out_vad=my_video_fast.mp4

# 3. Multi-angle: detect on one file, speed up all files in a folder
just speed-all folder=recordings/ master=recordings/facecam.mp4 folder_output=recordings/output/

# 4. Browse output files in your browser (requires Docker)
just browse browse_dir=recordings/output/

Run just with no args to see all available recipes.

• Python 3.x
• FFmpeg & FFprobe: Must be installed and accessible in your system's PATH.
• Python Packages: Listed in videospeeder_project/requirements.txt. Install them via pip.

  pip install -r videospeeder_project/requirements.txt

  (Note: --vad requires installing PyTorch + Silero VAD; this can add significant install size.)
• (Optional) NVIDIA GPU & Drivers: For GPU acceleration features.
• (Optional) CUDA Toolkit: Often needed for Whisper's GPU support during transcription.

1. Clone the repository:

   git clone https://github.com/jakkaj/videospeeder.git
   cd videospeeder

2. Install Python dependencies:

   # Navigate into the project directory if you aren't already
   cd videospeeder_project
   pip install -r requirements.txt
   # Or use the Makefile shortcut from the project root
   # make install # (Run this from within videospeeder_project/)

   (Note: The make install command in the Makefile assumes you are inside the videospeeder_project directory).

VideoSpeeder consists of two main scripts located in the videospeeder_project directory.

Run this script from within the videospeeder_project directory.

python videospeeder.py --input <your_video.mp4> --output <output_video.mp4> [OPTIONS]

Common Options:

• -i, --input: Path to your input video file (Required).
• -o, --output: Path for the processed output video file (Required).
• -t, --threshold: Silence threshold in dB (Default: -30.0). Lower values detect quieter sounds as silence.
• -d, --duration: Minimum duration of silence in seconds to be sped up (Default: 2.0).
• --no-vad: Disable VAD and fall back to FFmpeg silencedetect (amplitude-based).
• --vad-threshold: Speech probability threshold in [0.0, 1.0] (Default: 0.75). Higher rejects more noise.
• --indicator: Show the >> [Speed]x overlay during sped-up parts.
• --gpu: Enable NVIDIA NVENC GPU encoding.
• --gpu-decode: Enable NVIDIA CUVID/NVDEC GPU decoding (Experimental).
• --offset: Start processing from this time (in seconds).
• --process-duration: Process only this duration (in seconds) from the start (or offset).
• --detect: Detect speech/silence and write a .vad.json sidecar file next to the input video, then exit. No output video is produced.
• --vad-json: Load silence intervals from a previously generated .vad.json sidecar instead of running live detection.
• --quiet: Suppress informational output (progress bars and summary still shown).
• --folder: Process all video files in a directory using a shared .vad.json sidecar.
• --vad-master: Master video file for VAD detection in folder mode. Detects speech on this file, writes a sidecar, then processes all videos in --folder.
• --overwrite: Re-process videos even if output files already exist (default: skip existing).
• --extensions: Comma-separated video file extensions for folder mode (Default: mp4,mkv,mov,avi,webm).
• --parallel N: Process N videos simultaneously in folder mode (Default: 1). Best with --gpu; each video uses one NVENC session.

Example:

# Process a video (VAD speech detection is on by default)
python videospeeder.py -i my_recording.mp4 -o my_recording_fast.mp4

# With GPU encoding and speed indicator overlay
python videospeeder.py -i my_recording.mp4 -o my_recording_fast.mp4 --gpu --indicator

# Tune VAD sensitivity for very noisy audio
python videospeeder.py -i my_recording.mp4 -o my_recording_fast.mp4 --vad-threshold 0.80

# Fall back to amplitude-based detection (no VAD)
python videospeeder.py -i my_recording.mp4 -o my_recording_fast.mp4 --no-vad -t -40

If you record with multiple camera angles (e.g. facecam, screen capture, overhead), VideoSpeeder can process them all with identical timing from a single speech-detection pass.

Two-step workflow (detect once, then process all):

# Step 1: Detect speech on the master audio source (writes facecam.vad.json)
python videospeeder.py -i recordings/facecam.mp4 --detect

# Step 2: Process all videos in the folder using the shared sidecar
python videospeeder.py --folder recordings/ -o recordings/output/ --gpu

One-liner workflow (detect + process in a single command):

python videospeeder.py --folder recordings/ --vad-master facecam.mp4 -o recordings/output/ --gpu

Parallel processing (process 2 videos at once for ~1.8x faster throughput):

python videospeeder.py --folder recordings/ -o recordings/output/ --gpu --parallel 2

Both workflows produce output files with identical segment timing, so they stay in perfect sync when imported into your NLE.

Folder mode features:

• Auto-discovers the .vad.json sidecar in the folder (errors if zero or multiple found)
• Skips already-processed outputs by default; use --overwrite to re-process
• Continues processing if a single file fails, with a summary at the end
• Handles duration mismatches between angles (truncates silence intervals per video)
• --parallel N processes N videos simultaneously; --parallel 2 is the sweet spot for GPU encoding. Consumer GPUs support 8-12 concurrent NVENC sessions

Offline note for --vad:

• Silero VAD may download/cache model assets on first use. If you're offline, run once while online or preload the model in your environment before using --vad without network.

Run this script from within the videospeeder_project directory.

python transcribe.py --input <your_video_or_audio.mp4> --output <subtitles.vtt> [OPTIONS]

Common Options:

• -i, --input: Path to your input video or audio file (Required).
• -o, --output: Path for the output VTT subtitle file (Required).
• -m, --model: Whisper model size (Choices: tiny, base, small, medium, large. Default: large). Larger models are more accurate but require more resources.

Example:

# Transcribe a video using the 'medium' Whisper model
python transcribe.py -i my_recording_fast.mp4 -o my_recording_subs.vtt -m medium

The Makefile inside videospeeder_project provides shortcuts for common tasks (run from within videospeeder_project directory):

• make install: Install dependencies.
• make test: Run a test using a sample file path (edit the path in the Makefile first!) with GPU and indicator.
• make test-segment: Run a test on a segment.
• make transcribe INPUT=video.mp4 OUTPUT=subs.vtt [MODEL=large]: Transcribe a file.
• make transcript-segment: Transcribe the output of make test-segment.
• make clean: Remove generated .mp4 files from the directory.
• make help: Show available make commands.

This project is licensed under the MIT License. See the LICENSE file for details.

Found a bug or have an idea? Feel free to open an issue on the GitHub repository!