Learning voice conversion on a 4 GB GPU
Jul 2023 - Jan 2024My first real exposure to machine learning came from trying to make AI song covers with Retrieval-based Voice Conversion. I did not understand much of the theory yet. I was preparing audio, running training for 1,200 to 2,000 epochs, and adjusting CUDA settings on a laptop with 4 GB of VRAM.
Original Song Credits: Licensed to YouTube by T-Series (on behalf of composer Pritam & lyricist Rajat Arora; original artist Javed Bashir); and Indian music rights societies.
Original Song Credits: Licensed to YouTube by T-Series (on behalf of composer/lyricist Vishal Dadlani & music director duo Vishal-Shekhar; original artist KK); and Indian music rights societies.
I did not know about Google Colab at the time, so I kept working within the laptop’s limits. Preparing the data, waiting for a run, and listening to the result made the process unusually tangible. I could hear when something had gone wrong, even when I could not yet explain it properly.
The samples here are from those experiments.
