Tool comparison

DrumSplit vs Demucs

By DrumSplit · Updated 4 Oct 2026

Demucs and DrumSplit solve different parts of drum separation. Demucs can run on your computer, and its standard output is a single drums stem alongside bass, other and vocals. DrumSplit is a browser service that returns 7 files: kick, snare, hi-hats, toms, cymbals, a drum bus and the music without drums. Choose based on the files you need and where you want the audio processed.

At a glance

How to run Demucs yourself

Demucs is an open-source Python and PyTorch tool. In a terminal, install it with:

Then run it on an audio file:

On the first run, Demucs downloads its model weights automatically. By default, the selected model is htdemucs. When processing finishes, look in separated/htdemucs/song/ for drums.wav, bass.wav, other.wav, and vocals.wav. The folder name uses the track name, so it will differ if your input file has a different name.

CPU processing works, but it is much slower than using a CUDA GPU. Choose the device explicitly with -d cpu or -d cuda, for example demucs -d cpu song.mp3. CUDA is the option for a compatible NVIDIA GPU and installed CUDA setup. If you want Demucs to return only drums and the rest of the mix, run demucs --two-stems=drums song.mp3. That creates drums.wav and no_drums.wav.

There are model variants. -n htdemucs_ft selects the fine-tuned model, which is described as better but takes roughly four times longer. -n htdemucs_6s adds guitar and piano stems. For example, demucs -n htdemucs_6s song.mp3 selects that model. Start with the default if you simply want to see Demucs's standard four outputs; change the model when those additional outputs or the fine-tuned option fit your job.

When Demucs alone is enough

If you need one drum track to rebalance against the rest of a song, Demucs may be all you need. Its default run also returns bass, other, and vocals, which can be useful when your project calls for those parts. Use --two-stems=drums when you only want the drum stem and the no-drums version.

DrumSplit has a narrower output. Natural mode returns 7 files: kick, snare, hi-hats, toms, cymbals, a drum bus and the music without drums. Direct mode returns the five kit pieces when the input is already drums-only; it skips isolating drums from the full mix. DrumSplit does not produce separate vocal, bass, or guitar stems.

When you need kit pieces

The standard drums.wav still contains the whole kit together. That is the key decision point: if you need to adjust or arrange kick, snare, hi-hats, toms and cymbals independently, a single drums stem is where you start rather than where you finish.

DrumSplit separates those five kit pieces from an uploaded song and adds a drum bus and the music without drums. Natural is the general default for acoustic separation across genres. Puncher adds transient enhancement and is tuned for EDM, hip-hop, pop, and heavily compressed mixes. The output is 7 files as 24-bit WAV in a ZIP. Direct mode is for an input that already contains drums only, such as a drum bus bounce, loop, break, or a drum stem made with another tool.

The two-step recipe

1. Run demucs --two-stems=drums song.mp3 and locate the generated drums.wav in the output folder.
2. Open DrumSplit and upload that drum-only file using Direct mode.
3. Download the ZIP containing the separate kick, snare, hi-hats, toms, and cymbals.

This workflow uses Demucs to isolate the drums from a full song, then DrumSplit to divide that result into kit pieces. It adds a second processing step, a separate DrumSplit charge, and another model pass over audio that Demucs has already processed. Use it when the individual kit files are worth those costs. If a single drum track is enough, stop after Demucs.

Privacy and where processing happens

Demucs processes audio locally on the computer where you run it. That means setup, model download, and the processing workload belong to your machine. DrumSplit processes uploaded audio in the cloud. Results remain in My Splits for 24 hours and are also emailed; uploaded audio is deleted after 24 hours.

Pick the workflow that fits the recording and your setup. For material you need to keep on your own computer during separation, run Demucs locally. If you use DrumSplit, upload only audio you have the right to use and are comfortable sending for cloud processing. You must have rights to the original audio to use the resulting stems commercially.

Cost and time

Demucs is free under the MIT license, but free software still has setup and hardware costs in practice: install Python and PyTorch, download model weights, and let your computer do the work. CPU is supported and much slower than GPU processing. The fine-tuned htdemucs_ft model takes roughly four times longer than the default model.

DrumSplit costs from $0.99 per split (10 for $4.99, 50 for $14.99, 150 for $29.99), with no subscription and no free tier. Purchased credits never expire. A typical three-to-four-minute song takes about two to three minutes to process. It runs in a modern browser and needs no installation or user-side GPU. Uploads can be up to 100 MB and eight minutes long, whichever limit is reached first; 5.1 and other multichannel files are not supported, so convert those to stereo first.

Common questions

Does Demucs separate kick and snare?

The standard Demucs output is a single drums stem, with bass, other and vocals alongside it. DrumSplit returns 7 files: kick, snare, hi-hats, toms, cymbals, a drum bus and the music without drums.

Can DrumSplit replace Demucs for vocals or bass?

No. DrumSplit is focused on drums. Demucs can return vocals, bass, and other stems in its default four-stem mode.

Can I send a Demucs drum stem to DrumSplit?

Yes. Upload the Demucs drums.wav output in Direct mode, which is for audio that already contains drums only.

Do I need a GPU to run Demucs?

No. You can select CPU with -d cpu, though it is much slower than a GPU. A CUDA GPU can be selected with -d cuda.

How long and how large can a DrumSplit upload be?

The file must be no more than 100 MB and no longer than eight minutes. Whichever limit you reach first applies.