Spleeter

Deezer's audio team released Spleeter to split finished mixes back into their component stems, and it has become a de facto baseline for music source separation. Pretrained TensorFlow models ship in three configurations: 2 stems (vocals and accompaniment), 4 stems (adding drums and bass), and 5 stems (adding piano), with the 2 and 4 stem models reported to perform well on the musdb benchmark. On a GPU the 4-stem model separates audio around 100 times faster than real time, and CPU-only operation works for smaller jobs. Installation is a pip, conda, or Docker pull away, with a simple command-line interface plus a Python API, and a Google Colab notebook covers experimentation without local setup. Release 2.4.2 from April 2025 is the current PyPI version and supports Python 3.8 through 3.11, the code is MIT licensed so commercial use is fine, and the repository has around 28,300 stars. Producers, karaoke services, DJs, and music information retrieval researchers use it for remixing, backing-track creation, and building training datasets, though Apple Silicon users face TensorFlow compatibility friction.