MusicCaps

MusicCaps is a dataset composed of 5.5k music-text pairs, with rich text descriptions provided by human experts. For each 10-second music clip, MusicCaps provides:

1. A free-text caption consisting of four sentences on average, describing the music and

2. A list of music aspects, describing genre, mood, tempo, singer voices, instrumentation, dissonances, rhythm, etc.

Usage

conda create --name MusicCap python=3.9

conda activate MusicCap

pip install datasets yt-dlp pydub

Install FFmpeg

For Windows: Download FFmpeg, add the path to the system's environment variables.

For macOS/Linux:

# macOS
brew install ffmpeg

# Ubuntu/Debian
sudo apt update
sudo apt install ffmpeg

Pass cookies to yt-dlp

• Log in YouTube
• Use a conforming browser extension to export cookies, such as Get cookies.txt LOCALLY and Cookie-Editor for Chrome, cookies.txt for Firefox.
• Copy and save the cookie (Netscape format) to the local file cookies.txt

Run

python Download.py