Open source made stem separation common but not a rights grant

What happened

Music source separation, splitting a mix back into stems such as vocals, drums and bass, moved from a hard research problem to a downloadable tool within weeks in late 2019. Deezer published Spleeter, and the project's release notes mark v1.4.0, dated 28 October 2019, as the 'first public release', shipping three pretrained configurations, two, four and five stems. A month later, on 27 November 2019, Meta's team submitted a competing paper introducing Demucs, a waveform-domain model reporting 6.3 dB signal-to-distortion ratio on the MUSDB benchmark. Both projects kept developing, notably in a 2022 paper adding a hybrid transformer architecture to Demucs, reaching 9.20 dB SDR with an enlarged training set.

What the documents say

The two tools take different routes to the same task: Spleeter processes a mix as a spectrogram using TensorFlow, while Demucs's original paper argues for modelling the raw waveform directly, reporting better perceived naturalness at the cost of more audible blending between sources. The 2022 Demucs paper states its strongest result required not just the standard MUSDB set but 'an extra training dataset of 800 songs', a concrete detail about what improves quality that a user comparing tools by ear would not otherwise know. Neither repository claims ownership over separated stems; Spleeter's repository instead carries a disclaimer instructing users to 'get proper authorization from right owners beforehand' when separating copyrighted material, a caveat Demucs does not carry in comparable terms.

Why it matters for makers

The mechanism worth separating from the marketing language is this: separation is not licensing. A tool producing a clean a cappella or an isolated drum stem from a commercial recording has not granted any right to use that stem - Spleeter's documentation says as much directly, and the absence of an equivalent line in Demucs's materials does not mean the legal position differs, only that it is unstated. For a producer building a remix or a sample-based track from separated stems, the source recording's rights do not disappear because a neural network did the separating.

What to check before you use it

Before using a separated stem in released work, check whether the source recording is one the maker has rights to separate and reuse, since neither tool's documentation substitutes for that clearance. Check which model version and training configuration produced a stem when quality matters, since the 2022 hybrid transformer result depended on training data beyond the standard MUSDB benchmark and is not representative of every configuration available. This is an editorial checklist, not a claim either project makes about downstream rights.

• Does the maker hold, or have a licence to, the rights in the original recording being separated?
• Which version and training configuration of the separation model was used, and does that match the quality being assumed?
• Is a separated stem being treated as newly owned material, when it remains derived from an existing recording?

Spleeter and Demucs are best remembered together as the moment stem separation left specialist studios for anyone with a laptop - an opening that neither project's documentation pretends settles who is allowed to pull a recording apart in the first place.

Continue in Listening & research, or Benchmark stem separation fairly: score the signal and hear the job.