Stem Splitter

Almost every online stem splitter works the same way: you send your track to a company's server, it runs a model there, and you download the result. For a song off an album that is merely a privacy question. For the files people most often want to split, it is a bigger one — an unreleased demo, a mix that is not finished, a client's session, a recording of a rehearsal or a service. Those are exactly the files that should not be sitting on a stranger's disk.

Here the model comes to your audio instead. The network is downloaded to your browser once and executed by your own graphics card or processor. Your file is decoded inside the tab, separated inside the tab, and the stems are written out inside the tab. Traffic goes one way only: weights come down, the song never goes up.

It is worth being exact about it. The model files are fetched from Hugging Face and the runtime that executes them from jsDelivr, so both see an IP address, the way any website does. What is never transmitted is the audio, and no server here ever holds it — so there is nothing to log, retain, train on, or leak.

The trade-off is time, and it is honest to say so. A data centre GPU separates a song in seconds; your laptop takes longer, and much longer without WebGPU. What you get in exchange is a tool with no queue, no account, no daily limit, and no question of whether you trust someone with the file.

The model is HT-Demucs, a hybrid transformer published by Meta's research team. "Hybrid" is the interesting part. Audio can be looked at two ways: as a waveform, the raw pressure wave over time, or as a spectrogram, a picture of which frequencies are sounding at each moment. Drums are easy to see in a waveform, where every hit is a sharp spike. A sustained vocal is easier to see in a spectrogram, where its harmonics form stacked lines. HT-Demucs reads the music both ways at once, lets a transformer compare the two views, and produces each stem from both branches added together.

A song is too long to look at in one go, so it is cut into windows of 7.8 seconds — the length the model was trained on — that overlap by a quarter. The model is least sure of itself at the edges of each window, where it has no context, so the windows are crossfaded into each other with ramps that add up to exactly one. The seams are inaudible, and a steady note passes through at exactly the level it went in.

Each window then becomes four stereo signals — vocals, drums, bass, and everything else — or six, if you chose the model that also pulls out guitar and piano. Because the model was trained so that its outputs add back up to its input, the stems are not separate guesses — they are one decomposition of the song. That is why turning every stem on gives you the original back, and why muting one gives you the song without it rather than something reconstructed.

On a browser with WebGPU, the model runs on your graphics card — the four-stem model in 21 short pieces, so the page stays responsive while it works. Without it, the same model runs on the processor through WebAssembly. Both produce the same stems; one is simply several times faster. If the graphics card turns out not to cope — some do not — the tool notices and switches to the processor without losing the work already done.

Practising an instrument. Mute the part you play and play it yourself with the real band. Bassists mute the bass, drummers mute the drums, singers mute the vocal. Solo the part instead and you can hear exactly what is being played, without the rest of the arrangement on top of it — which is the fastest way to learn a line by ear.

Karaoke and backing tracks. "No vocals" is an instrumental built from the record itself, not a cover version. For worship teams and cover bands, the same idea makes a rehearsal track: drop the part that a real person will be covering live.

Remixing and sampling. An a cappella to put over a new beat, a drum break on its own, a bassline to chop. These are what producers used to need official stems or a lucky multitrack leak for.

Transcription and study. Hearing the bass clearly is most of working out a song's harmony; hearing the drums alone is how you write out a groove. Teachers use stems to show a class what each instrument is contributing to a mix.

Rescuing a recording. A rehearsal where the drums swamp the singer, a phone video of a performance, an old demo with no multitrack: separating it lets you rebalance what was captured as one stereo file.

Start from the best file you have. The model can only separate what is there. A lossless file or a high-bitrate MP3 gives cleaner stems than a 128 kbps copy or the audio ripped from a video, because compression smears exactly the fine detail the model uses to tell instruments apart.

Expect "other" to be the messiest stem. It is everything left over, so it holds whatever the model was not sure about. If you need guitar or piano on its own, split with six stems — and with four, anything you want from "other" is a starting point, not the finished article.

Use volume rather than mute for practice. Taking a part down to around 20% instead of switching it off keeps a guide you can lean on while learning, and hides most of the faint bleed between stems.

Mind the clipping on a custom mix. Pushing several stems above 100% can make the mix louder than the original. The WAV is clamped rather than wrapped, so it distorts instead of exploding, but it is better to pull the others down than to push one up.

Live recordings and old mono records are harder. The model learned from modern studio productions. Heavy reverb, a crowd, or a mono mix with everything in the same place will separate less cleanly — still useful, but with more bleed.