Change Your Singer: A Transfer Learning Generative Adversarial Framework for Song to Song Conversion

I trained a model that lets you listen to any song you want in the voice of any artist you want using CycleGANs from images but applied to audio.

This was part of my undergrad research internship and my first exposure to deep learning.

In this project:

1. Developed an end-to-end deep learning pipeline for song-to-song conversion powered by GANs
2. Collected & labelled non-parallel audio datasets of singers from YouTube
3. Used pre-trained U-Net to extract vocals
4. Trained CycleGAN to convert vocals to target singers in TensorFlow