deezer/spleeter ? reverse-engineered prompt

Build me a Python music separation tool that can take an audio file and split it into stems like vocals and accompaniment, and also support the fuller options with drums, bass, piano, and other. I want it to work from the command line for quick use, and also be usable as a Python library in my own scripts.

Please include pretrained models so I can run it right away on a song without training first, but make it possible to train on isolated source datasets too if needed. It should save the separated tracks as audio files in an output folder and be fast enough to use on longer songs without waiting forever.

If there are any setup steps that are needed for audio handling, make them easy to follow, and if you need to check current docs or best practices online, go ahead. Keep the project clean, tested, and ready to run from the repo.