Demucs-Family Stem Separation

An open-source separation family that splits vocals, drums, bass and other offline, in batches.

Model profile

Vendor: Meta（Facebook Research）
Form: 本地模型
Input: 混音音频
Output: 分轨
Licence: 代码许可与预训练权重许可要分开核对，训练素材的授权范围以官方仓库与数据集说明为准
Verified: 2026-10-30

What it lets you finish

Batch-separate a folder of mixes into vocals, drums, bass and other; make an instrumental draft of an old song; cut a usable drum or bass fragment for sampling; prepare reference material for training another model; process a master that may not leave the machine.

Inputs and outputs

You input mixed audio files and get one wav per stem, with sample rate and bit depth set by you. Internally the model resamples to its own working rate and back, so on long files watch the start alignment and latency. Both a command line and a Python interface exist, and whole directories can be processed; a GPU is fast, a CPU is slow but finishes.

Strengths

Fully offline, so unreleased masters and confidential client work stay on the machine.
Batchable and scriptable, with results tied to a weight set; reruns of the same input align.
Vocal and backing separation is steady among open-source options, enough for drafts and sampling.
Several stems per run, each written as its own file for per-track repair.
No per-minute billing; the cost is electricity and time.

Weaknesses and failure modes

Residue: with heavy reverb, distortion or live recordings, the vocal stem turns watery and the backing keeps a vocal shadow.
Low end: bass and kick share a band, so low end thins after separation or the two stems bleed into each other.
Drums: cymbal and snare transients round off, and hi-hat metal smears into noise.
Alignment: internal resampling shifts timing by milliseconds, so summing stems against the original cancels in phase unless you align first.
Resources: long pieces eat video memory, and a whole album on CPU takes a long wait; batching is your own scheduling problem.
Weight drift: results differ noticeably between weight sets, so swapping weights is swapping tools; record which set a project used.

Terms, price and alternatives

Questions to check: do the code licence and the pretrained weight licence agree? Does the public training dataset permit commercial use? Separated stems remain bound by the copyright of the original recording, so is that stated in your delivery notes? Is training your own model on the stems allowed?

Price: the software is free and the cost is hardware and time; hosted versions run by others are billed per subscription or per use, as stated on their official pages.

Alternatives: LALAL.AI when you do not want to install anything and need a one-click web result; a GUI tool in the UVR family when you want to try many models without commands; iZotope RX when separation and repair should happen in one place inside the DAW.

Listen in the library

little%20town%20 %20orchestral%20midiC13:12A crowded game orchestral score with many parts sharing bands, among the hardest material to separate; a reference for this kind of output, not a generation by this model (C1)DetailsLibertangoC14:42Accordion and chamber parts packed into the midrange, to hear which stem thins; a reference for this kind of output, not a generation by this model (C1)DetailsChorale BWV 1/6C11:02A four-part chorale whose voices differ only by register, where bleed is easiest to hear; a reference for this kind of output, not a generation by this model (C1)Detailssoul-groove9 · 105 BPM · beat 4/4C10:58A soul kit where kick and bass share the low band, for checking the low end after separation; a reference for this kind of output, not a generation by this model (C1)Details

All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).

Sources

Demucs 的模型结构与推理方式，依公开论文摘要与仓库说明核对，未抄原文
分离残渣与低频损失的处理，依混音实践通识核对，未抄文本

Used for fact-checking only; all prose is original.