Stem models
Cloud vs Local Stem Separation: Which to Use?
The first wall most people hit with music source separation is not the upload button. It is the menu. Open-source projects such as Demucs, Ultimate Vocal Remover (UVR), Spleeter and Open-Unmix all promise clean vocals, drums or bass, and they publish benchmark numbers that are hard to compare at a glance. Those names are widely used technical labels in research and hobby tools. The comparison page presents the choice as Cloud Standard versus Local 6 Tracks rather than under those project names.
What the site does offer is a practical choice between two modes: Cloud Standard versus Local 6 Tracks. Cloud Standard splits a song into two or four tracks on remote servers. Local 6 Tracks runs on your own computer and returns drums, bass, vocals, guitar, piano and the remaining material. Both are available from the same AI Stem Splitter, and Local 6 Tracks requires a sign-in. That splitter page states that Local 6 tracks runs the open-source HT-Demucs model (Demucs v4, htdemucs_6s) in your browser, and that Cloud Standard uses a different third-party model through WaveSpeed. SampLab Pro is an independent service and is not affiliated with Meta or the Demucs authors. The goal of this guide is a decision you can explain in one sentence: which mode fits the stems you need, the device you have, and whether the audio should leave your machine.
Process only files you own or have permission to use. The upload form asks you to confirm that; separating a commercial recording does not grant the right to publish karaoke tracks, acapellas or isolated stems.
Start From the Job, Not From a Famous Model Name
Selection gets easier if you answer two questions before you care about architectures.
Do you need two groups, or many? Karaoke-style instrumentals, acapella study and “voice versus the rest” only require a two-way split (vocals and accompaniment). Remixers, transcribers and anyone who wants a dedicated bass or drum file need a multi-way split. Industry datasets such as MUSDB18 popularized a four-stem grouping—vocals, drums, bass and “other”—which is why so many tools talk in those buckets. A six-way split that also estimates guitar and piano is a different, finer request.
What are you optimizing for? Speed and convenience on any browser point one way. Keeping the file on the device, or hearing guitar and piano as their own tracks, point another. Quality and speed are a trade-off in every published system; running every sketch at the heaviest setting is usually wasted effort.
SampLab maps those questions onto two modes rather than onto a desktop installer:
| If you need… | Prefer | Why |
|---|---|---|
| Vocals and accompaniment, or vocals / accompaniment / bass / drums, from any browser | Cloud Standard | Two or four tracks, processed on servers, not tied to your laptop’s GPU |
| Drums, bass, vocals, guitar, piano and the rest, with the audio staying on the device | Local 6 Tracks | Six parts estimated on your computer (sign-in required) |
| A quick listen to decide whether a song is even worth splitting | Cloud Standard | Works wherever you can upload a file |
| Sensitive material that should not leave the machine | Local 6 Tracks | The comparison page states that the audio never leaves the device |
If you only wanted an EQ-style low/mid/high split, you are in the wrong article. That is filtering, not source separation; the frequency band splitter makes that distinction explicit.
What Cloud Standard and Local 6 Tracks Actually Return
Cloud Standard separates a song into two or four tracks on SampLab’s servers. Two tracks give you vocals and accompaniment. Four tracks add bass and drums beside the vocal and the remaining accompaniment. The work is not limited by whether your laptop has a dedicated GPU, because the heavy lifting is remote. The practical cost of that convenience is the obvious one: the file has to be sent away to be processed. Use it when you are on a machine you do not want to load with models, or when two or four stems are enough.
Local 6 Tracks runs on your computer and returns drums, bass, vocals, guitar, piano and “other.” The comparison page notes that the vocal, drums and bass are the strongest of those six. The first use downloads about 147 MB of model files, and the job needs a few gigabytes of memory, so it is aimed at computers rather than a phone tab. In the site’s own test, a 2 minute 43 second song took about a minute with GPU acceleration and several minutes without it. Sign-in is required. After that download, the audio stays on the device.
Both modes live behind the same upload widget on the comparison page and the splitter: common audio formats, typically from 1 second to 6 minutes and up to 50 MB. Confirm the checkbox that you own the audio or have permission to process it. Preview, mute and solo before you download a single stem or a ZIP of several tracks; a clean intro is a weak test if the chorus is densely layered.
The same splitter can send a stem onward to WAV to MIDI or Audio to Sheet Music when you want notes instead of another audio file. For the conceptual background—bleed, artifacts and why stems are estimates—see the earlier guide to stem separation.
Industry Models You Will See Named Elsewhere
Writers and forums still compare four well-known open-source lines. The short version below is general knowledge about those research projects. For what SampLab itself runs, use the splitter-page description above: Local 6 tracks is HT-Demucs (Demucs v4, htdemucs_6s); Cloud Standard is a different third-party model through WaveSpeed. This article does not guess at any other checkpoint names.
- Demucs (Meta / open-source research lineage) is a hybrid approach that looks at waveform and spectrogram information. The paper Hybrid Spectrogram and Waveform Source Separation is a common entry point for that family. Public checkpoints are discussed in four-stem and six-stem variants (the extra pair usually being guitar and piano). Treat any quality ranking you see online as a result on a specific test set, not as a guarantee for your mix.
- Ultimate Vocal Remover is a desktop GUI that hosts many models rather than a single algorithm. People reach for it when they want to drag and drop a file and try several vocal/instrumental networks, including ones often grouped under names like MDX-Net.
- Spleeter (Deezer) shipped pretrained 2-, 4- and 5-stem models and became famous for speed—useful as a fast first pass when you have a large batch and can accept an older quality bar.
- Open-Unmix is a readable reference implementation used as a research baseline. Some of its weight packages use different licences than the code (UMXL is commonly flagged as non-commercial); that detail belongs to Open-Unmix, not to SampLab.
If your actual need is “I want vocals isolated in the browser,” start with Cloud Standard or with a scene page such as the acapella extractor. If your need is “I want the instrumental without the lead vocal,” the vocal remover is the matching entry point. Neither decision requires you to install those four projects.
Three Licence Layers That Are Easy to Mix Up
Commercial and even serious hobby projects trip over the same confusion: “it is open source, so I can ship the result.” There are at least three separate questions.
- Code licence. Many of the projects above distribute code under permissive terms such as MIT. That governs the software, not your song.
- Weight licence. Trained checkpoints can use a different licence from the repository. Authors sometimes restrict weights to research use, or ship a higher-quality pack under a non-commercial Creative Commons licence. Anyone who self-hosts a model should read the model card, not only the README badge. This article is not legal advice and does not interpret SampLab’s own licensing page.
- The rights in the audio you split. Even a fully permitted engine does not clear the composition or the sound recording. Pulling vocals or drums from a commercial track and distributing them still needs the rights holder’s permission.
SampLab’s form restates layer three as a checkbox. Treat it as a real legal step.
Do Not Choose by a Single SDR Number
Forum tables that say one model is “9.2 dB” and another is lower look decisive until you notice they were often measured on different datasets, versions and aggregation methods. SDR and related scores from tools such as museval (BSSEval v4) are useful for ordering systems under identical conditions. They are a poor substitute for listening to your material.
A practical habit, borrowed from how engineers actually pick a stack, is:
- Use published benchmarks only to short-list two or three approaches.
- Run the same songs you care about—your genre, your vocal style, your live versus studio recordings—through the candidates.
- Judge the hook and the busiest chorus, not only the intro, and listen for residue, missing transients and watery highs rather than staring at one number.
On SampLab, that comparison is local to the two modes on the page: process the same file with Cloud Standard and with Local 6 Tracks (when your computer can run the latter) and keep the version that serves the part you need. Official league tables will not tell you whether the guitar estimate on your particular song is usable.
Remote Processing Versus On-Device Processing
The older “build versus buy” debate in self-hosted tools is the same tension in simpler clothes. Running an engine yourself means you handle setup, memory and GPUs, and the files can stay in-house. Calling a remote service means someone else’s machines do the work, which is faster to start and requires the audio to travel.
Cloud Standard is the remote path: any browser that can upload a file, two or four stems, no 147 MB local download. Local 6 Tracks is the on-device path: sign in, download the model files once, spend CPU or GPU time, keep the song on the machine, and receive six estimates including guitar and piano. If you are on a shared or low-memory device, Cloud Standard is the realistic choice. If the recording should not leave the laptop, or you specifically want those extra instrument buckets, Local 6 Tracks is the one that matches the requirement.
There is no all-purpose winner. Requirement first (two-way or multi-way, which instruments), then constraint (device, privacy, whether you can sign in), then measurement on your own files.
Frequently Asked Questions
What is the difference between Cloud Standard and Local 6 Tracks? Cloud Standard returns two or four tracks from a server-side split. Local 6 Tracks estimates six parts on your computer—drums, bass, vocals, guitar, piano and other—after a one-time download of about 147 MB. Local 6 Tracks needs a sign-in, and the audio does not leave the device.
Does SampLab name Demucs, UVR or Spleeter on these pages?
The Cloud versus Local comparison page describes the two modes without those labels. The AI Stem Splitter page does say that Local 6 tracks runs open-source HT-Demucs (Demucs v4, htdemucs_6s) in the browser, and that Cloud Standard uses a different third-party model through WaveSpeed. It also states that SampLab Pro is not the official Demucs project and is not affiliated with Meta or the Demucs authors. UVR, Spleeter and Open-Unmix appear in this article only as systems you will encounter elsewhere.
Which mode should I use for karaoke or an acapella? A two-track vocal/accompaniment split is usually enough. Cloud Standard covers that without asking your laptop to load a local model. Scene-specific pages exist for extracting the vocal or removing it.
Which mode should I use if I need guitar or piano separately? Local 6 Tracks is the mode that lists guitar and piano among its six outputs. Preview those stems carefully; the site notes that vocals, drums and bass are the strongest of the six.
How long does Local 6 Tracks take? It depends on the machine. The published on-page test processed a 2 minute 43 second song in about a minute with GPU acceleration and in several minutes without it. The job also needs a few gigabytes of memory.
Can I split any song I find online? Only process audio you own or have permission to use. Technical separation is not a licence to publish or sell the stems.
