Services like Spotify can generate personalized playlists by leveraging data on what songs their users like. Can we do the same without leveraging user data? In particular, can we build playlists leveraging only song metadata and audio content?
VibeSync is a research project which aims to explore the boundary of ML research with music. Inspired by recent advances with contrastive learning and joint language-audio embeddings, we aim to build a proof-of-concept system where a user specifies a playlist title and receives recommended songs. We want to see how far take this and what insights we can gain.
To install uv and create a virtual environment just run:
pip install uv
uv sync -p 3.11.5To start the virtual environment, run:
source .venv/bin/activateTo start the app, run make app
To use our scraper, create a .env file in the root directory of the project and add the following environment variables (filled with valid credentials for Jamendo):
CHROME_USER=<put your email here>
CHROME_PASSWORD=<put your password here>
CHROME_USER_2=<put your email here>
CHROME_PASSWORD_2=<put your password here>
Add more users as is needed. Then run make scrape to run our scraper.
VibeSync was made possible by Michael Ngo, Nancy Chen, Koji Kimura, Samantha Vaca, and Takuma Osaka.
Reference: https://github.com/microsoft/CLAP