Spleeter

Spleeter is Deezer's music source separation library for developers and audio researchers who want to split recordings into separate vocal and instrumental tracks on their own hardware. It includes pretrained models, so you can separate audio without first training a model. The library is open source under the MIT license.

Its models produce different sets of stems, the separate audio tracks extracted from a mix:

• Vocals and accompaniment for isolating a singing voice from the backing music.
• Vocals, drums, bass and other sounds for separating the main parts of a recording.
• Vocals, drums, bass, piano and other sounds when you need a separate piano track.

Written in Python with TensorFlow, Spleeter works as a command-line tool or as a library within an audio processing application. It also supports Docker. This makes it a fit for people building their own separation tools or processing recordings through an existing development pipeline.

GPU processing is a notable strength. Deezer reports that four-stem separation can run at 100 times real-time speed on a GPU. Researchers can also train their own separation models if they have a dataset of isolated sound sources.

Spleeter's pretrained models have been used in audio software including iZotope RX, SpectralLayers, Acoustica and VirtualDJ. The Spleeter 4 Max project brings it into Ableton Live. Deezer also offers a separate commercial product, Spleeter Pro, with professional support.