spleeter

Why we included this project

Anyone who works with recorded music eventually runs into the problem of wanting the vocals without the backing track, or the drums pulled out on their own. Spleeter, Deezer's source separation library, does this with pretrained TensorFlow models, from a simple vocals-versus-accompaniment split up to five stems that also isolate drums, bass, and piano. It is fast enough that a GPU can process a track far quicker than real time, and you can use it from the command line, as a Python library inside your own pipeline, or through Docker, so it slots into batch jobs or a web service without much fuss. It also supports training your own separation model if you have isolated source audio, which makes it useful as a ready-made tool and a base for custom work. The two- and four-stem models score well on the musdb benchmark, a reasonable sign that the output quality holds up.