Split a song into a percussive stem and a tonal stem, and download both as WAV files.
It separates a mix into two stems: the hits and the notes. Mute the drums stem and you have a bed to program new drums under. Keep only the drums and you have a break to chop. Rebalance the two and you have a remix that still sounds like the record.
This is harmonic and percussive source separation, the Fitzgerald median filtering method, and the idea behind it is visual. Look at a spectrogram of a song. Sustained notes draw horizontal lines, because a pitch holds one frequency over time. Drum hits draw vertical lines, because a strike covers every frequency for a moment. Those two shapes can be told apart by filtering.
Because the masks are complementary, the two stems reconstruct the input: nothing is invented and nothing lost, the energy is only routed.
The rule it applies is transient against sustained, which is not the same as drums against instruments. A plucked string starts with a transient, so guitar and piano attacks leak into the drums stem and can leave the music stem sounding soft. Cymbal washes and reverb tails are sustained, so they leak the other way. Tuned percussion is genuinely both. Vocals stay in the music stem, since a voice is harmonic, though the consonants often do not. Very short files, under about 17 analysis frames, come back unchanged: not enough spectrogram to filter.
No. The file is decoded in this tab and separated in a Web Worker on your own machine, and the WAV files are encoded locally. Nothing is sent to a server and there is no account. It works with the network switched off.
Because it is doing real work. Every frame is transformed to a spectrum, median filtered along two axes, masked and transformed back, once per channel. Expect roughly the length of the song again in processing time on a laptop. It runs in a Web Worker, so the page stays responsive and the progress bar is honest rather than decorative.
Because a sung note is harmonic, not percussive. This separation splits sustained tonal material from short broadband hits, so it cannot tell a voice from a guitar. If you want the voice out, use the vocal remover instead, which cancels the center of the stereo field. The two tools answer different questions.
Consonants, pick noise, piano hammers and other transients of tonal instruments read as percussive and leak into the drums stem. Ride cymbals and long crashes are sustained enough to leak the other way. Tuned percussion such as timpani or a marimba is genuinely both, so it splits across the two. Heavy distortion smears everything and blurs the distinction.
Not from this method. Separation into vocals, bass, drums and other needs a trained neural model such as Demucs, which needs a GPU server, which means uploading your song. This tool trades that ceiling for privacy and speed: deterministic arithmetic on your own machine that never sees a server.
Very nearly. The masks are complementary, they sum to one at every point in the spectrogram, so playing both stems together reconstructs the input apart from small overlap add rounding. That is what makes the split usable in a mix: you can rebalance the drums against everything else without a phase hole opening up.
Gusta Music is the free browser studio these tools came out of: multitrack timeline, sampling, mic recording, and an AI copilot that edits the song with you instead of generating one for you. Your work stays yours.
Open the studio