Test Drive: MusicGen via MLX
Published on .
Today I’m test driving MusicGen ported to MLX, for blazing fast generation on Apple Silicon. I do not fully understand MusicGen, but I do find the process of tokenizing audio fascinating. It uses parallel code books, which I think means layering of audio tokens to produce complex sounds. The process is extraordinarily simple:
git clone https://github.com/ml-explore/mlx-examples- Invoke
python -m generate
It takes just over 60 seconds on my M1 to generate 9 seconds of music with 500 steps. Here is an example.