Deezer's open-source Spleeter made stem separation instant with pretrained models

Splitting a finished song into its stems — vocals and accompaniment, or vocals, drums, bass, piano and the rest — is what karaoke, remix, DJ and audio-repair tools need, yet the individual parts exist only in a studio's multitrack files. Working from a mixed recording alone, source separation traditionally meant training a model for each job: assembling datasets of isolated sources, tuning a network, and running slow inference.

Deezer Research's answer, which surfaced publicly in late 2019, was Spleeter: a Python and TensorFlow library that ships already-trained, state-of-the-art models for 2-, 4- and 5-stem separation. It is engineered for throughput rather than bespoke accuracy — a GPU separates the four-stem version 100x faster than real time — and it installs with pip or Docker and runs from a single command, so even a whole catalog can pass through it quickly.

Because the hard part, the trained weights, came with the code under a permissive MIT license, third parties built on it instead of building their own: iZotope used the models for RX 8's Music Rebalance, Steinberg for SpectralLayers' Unmix, VirtualDJ for stem isolation and Algoriddim for NeuralMix, alongside community GUIs and websites. Research adopted it too — Spleeter became a baseline in the ISMIR 2021 Music Demixing Challenge — and the tool was written up in the Journal of Open Source Software.

Ship pretrained models; make separation 100x real-time

The trained weights ship with the tool, so users skip the hardest part of source separation — assembling isolated-source data and training a model — and get state-of-the-art results on the first run.

A one-line CLI with pip and Docker installs means a product team can embed separation in an afternoon instead of building it over months.

100x real-time GPU throughput turns the library into a pipeline primitive: processing an entire music catalog becomes cheap enough for streaming-scale features.

A permissive MIT license lets commercial vendors (iZotope, Steinberg, VirtualDJ, Algoriddim) put the models inside paid products, spreading Spleeter without Deezer spending on marketing.

When the scarce asset is a trained model, release it with the weights and make inference absurdly fast — the ecosystem adopts your tool as its baseline, and value moves to service and support

The release reached the front page of Hacker News on 3 November 2019 with 1,460 points and 175 comments. By 2026 the repository had grown to about 28,400 stars and 3,100 forks, and Deezer packaged the capability commercially as Spleeter Pro, offering finer separation, faster processing and dedicated support through Deezer Tech Services. The same pretrained models stayed in use across professional audio software and remained a reference baseline in the Music Demixing Challenge.