Autoclip

Turn a long YouTube video (or local file) into short, captioned, vertical clips — fully local, no paid APIs required. Good fit for faceless-video channels: no face tracking, just a blurred-background 9:16 reframe + clean word-by-word captions.

How it works

1. Download the source video with yt-dlp (or point it at a local file).
2. Transcribe locally with faster-whisper, getting word-level timestamps.
3. Score the transcript with text heuristics (hooks, questions, numbers, superlatives, filler-word penalty) to find the most clip-worthy moments — no LLM API needed.
4. Cut & caption: ffmpeg trims each clip, reframes it to a blurred- background 9:16 vertical layout, and burns in animated word-by-word captions.

Setup

# 1. System dependency: ffmpeg (not installable via pip)
#    macOS:   brew install ffmpeg
#    Ubuntu:  sudo apt install ffmpeg
#    Windows: choco install ffmpeg   (or download from ffmpeg.org)

# 2. Python dependencies
pip install -r requirements.txt

First run will download the Whisper model (a few hundred MB, one-time).

Usage

# From a YouTube URL
python main.py --url "https://www.youtube.com/watch?v=XXXXXXXX" --num-clips 5

# From a file you already have
python main.py --file /path/to/podcast.mp4 --num-clips 8

# Keep the original 16:9 aspect ratio instead of vertical reframe
python main.py --url "..." --horizontal

Output clips land in ./output/clips/clip_01.mp4, clip_02.mp4, etc.

Tuning

Everything lives in config.py:

• WHISPER_MODEL — bigger = more accurate, slower (base → small.en → medium.en)
• MIN_CLIP_DURATION / MAX_CLIP_DURATION — target clip length
• CAPTION_MAX_WORDS_ON_SCREEN — how many words shown per caption chunk
• Caption colors/fonts

Making highlight detection smarter (optional, still free/local)

The heuristic scorer in highlights.py works with zero setup, but it's not as smart as an LLM. If you install Ollama and pull a model (ollama pull llama3.1), you can swap in rank_with_ollama() from highlights.py to have a local LLM score each candidate clip instead — still completely free, just needs a decent GPU/CPU to run at reasonable speed.

Notes

• Whisper runs on CPU by default (compute_type="int8") so it works everywhere, but it's slow on long videos. If you have an NVIDIA GPU with CUDA, edit transcribe.py to use device="cuda", compute_type="float16" for a large speedup.
• This tool doesn't do face tracking/cropping — it's built for faceless content (podcasts, voiceovers, screen recordings) where a centered, blurred-background vertical layout looks clean without needing to track a speaker's face.