toolgarden.xyz
中文

Stem splitter

Free online stem separation that loads the open-source HT-Demucs model in your browser and splits a song into vocals, drums, bass, guitar, piano, and other

Upload a song

Choose a model

Choose stems to export

4 stems selected

Pick an audio file first.

Each model is downloaded the first time you use it (open-source HT-Demucs, MIT licensed; sizes shown above) over 4 parallel connections with a progress bar, then cached in your browser so later runs skip the download. The two models are cached separately.

Your audio is never uploaded. Decoding and separation both run locally in your browser; only the model file comes from a public CDN.

Separated stems

About this tool

Stem separation uses machine learning to estimate components such as vocals and accompaniment from a mixed recording. It has no access to the original multitrack project and works backward from the final mix, so vocal reverb, instruments sharing frequencies, and stereo effects can remain across outputs.

The model usually loads on first use, and processing time and memory grow with duration and device limits. Results suit practice, draft remixes, and analysis, not studio source tracks, and they do not bypass music copyright or performance licensing.

How to use it

  1. Upload a strong source

    Prefer lossless or high-bitrate audio rather than a file that has been repeatedly compressed.

  2. Load and separate

    Keep the page open and monitor model and processing progress.

  3. Listen to each stem

    Check vocal, drum, and accompaniment leakage and decide whether professional source tracks are required.

Supported range and limits

Output stems
Six tracks: vocals, drums, bass, guitar, piano and other
Model
An open-source source-separation model running in-browser; the first run downloads the model files
Time required
All six stems need a full inference pass, so a song typically takes several minutes and can fail on low-end devices
Separation quality
Depends on the mix. Arrangements with clear instrument separation come out well; dense mixes and heavily compressed masters leave bleed and residue
What it cannot do
Recover the original multitrack recording; the output is an estimate derived from the finished mix
Where it runs
The model is downloaded, but inference on your audio happens in the browser and nothing is uploaded

When you would use it

  • Creating practice accompaniment

    Reduce or isolate a lead vocal to rehearse melody, harmony, or an instrument part.

  • Analyzing a mix

    Listen separately to estimated vocals and backing to understand arrangement and frequency overlap.

  • Making a backing track to practise with

    Pull the instrumental out of a finished song to sing along to, or isolate the vocal to study the performance.

What to know before you start

  • Output is an estimate, and heavy reverb or overlapping frequencies can create watery, hollow, or leaking artifacts.
  • Long files can require substantial memory; trim to a shorter section when performance is limited.
  • Possessing separated files does not grant rights to copy, publish, adapt, or use them commercially.

Related concepts

stem
A grouped audio component of a full mix, such as vocals or accompaniment.
source separation
Computational estimation of multiple underlying sound sources from one mixed signal.

Frequently asked questions

Will the stems have artefacts?
Always some. Separation estimates each source from a finished mix, and the denser the arrangement and heavier the mastering compression, the more residue and bleed you get. Drums and bass usually come out cleanest; piano and guitar interfere with each other most.
Can I get the original studio multitrack?
No. Those exist only in the producer's session files. What you get here is an approximation inferred from the two-channel mix.
Why does it take so long?
Six stems require a full model inference pass, which is far heavier than ordinary format conversion. A three- or four-minute song typically takes several minutes, and modest hardware may fail partway.

Related tools