Free AI stem splitter

What is a stem splitter?

A stem is one part of a finished song on its own: the vocal, the drums, the bass, everything else. Producers get stems from the session, when they bounce each group of tracks as its own file. When all you have is the final stereo file, a stem splitter, also called a stem separator or stem extractor, lets you split stems out of it by estimating those parts with a neural network.

Ours gives you four stems in Fast. Vocals is the lead and backing voices. Drums is the kit and the percussion. Bass is the low line, whether it is a bass guitar or a synth. Other is everything that is left: guitars, keys, pads, strings, brass and effects. In HQ the guitar and the piano come out as their own stems, so Other holds what remains. People also look for the same thing as a song splitter, a music separator or an audio separator. If you only need to separate vocals and instruments, the vocal remover and acapella pages do that in one step.

What producers do with stems

Remix and bootleg: keep the vocal, rebuild everything under it. Sampling: lift a drum break or a bass line without the rest of the track in the way. DJ sets: mute the vocals to blend two songs, or keep only the drums and bass for a long transition. Learning: solo the bass of a record you love to see what it really plays.

Stems are also useful when you play along. Mute the drums and you have a drum remover, mute the bass and you have a bass remover, so you can play that part yourself with the rest of the record. With HQ you can also mute just the guitar or just the piano and play that part yourself. You get the isolated parts without asking anyone for the session files.

AI stem separation, explained

Fast uses Demucs v4, the hybrid transformer from Meta's AI research team, published as open source. It looks at the song in two ways at once, as a waveform and as a spectrogram, and learns to predict each stem from what it hears. It was trained on multitrack recordings, where the stems are known, which is how it learned what a voice, a snare or a bass line look like inside a mix.

HQ uses BS-RoFormer, a newer design that splits the spectrogram into frequency bands and lets a transformer follow each band through time and across the bands. The model we run is a six-stem one that people who work on music separation share openly. On public benchmarks this kind of model scores higher than Demucs v4. The price is size and time.

We run both with ONNX Runtime Web, on your graphics card when the browser has WebGPU (recent Chrome and Edge), which HQ needs. Fast can also run on your processor, which is slower. The song is cut into overlapping windows, about eight seconds in Fast and four in HQ, and cross-faded, so you do not hear seams. Your browser downloads a model once and keeps it.

How good are the stems?

Good, and not perfect. Demucs v4 is a well-known open model that many tools build on, but separation research moves fast and newer models score higher on benchmarks, which is why we added HQ. In practice, a clean stem is very usable for remixing, practice and reference, and the harder cases show: reverb tails of a vocal can end up in the other stem, cymbals can leak into it, and a very dense mix leaves faint traces of one stem inside another.

In our tests the four Fast stems added back together stayed very close to the original song, so the split does not throw any sound away. HQ estimates every stem on its own, so its six stems add up to the song a little less exactly, and on a song with sung vocals its vocal stem was clearly cleaner between phrases than the Fast one. A stem from a finished song will never be as clean as the same stem from the original session, and we would rather tell you that than promise it.

Tips for cleaner stems

Start from the best file you have. A WAV, FLAC or high-bitrate file separates better than a low-bitrate MP3, because compression smears the details the model uses. Skip any extra limiting first, and do not drop a file that has already been through a stem splitter. If Fast leaves traces you do not like, run the same song again in HQ.

If you are sampling, listen to the stems together first: all of them should sound like the song. Then solo the one you need and clean it like any sample, with a gentle high-pass on drums or a short fade at the cut. When a stem is almost right, layer a played part of your own over it and the leftovers disappear.

Free, private and in beta

Splitting happens on your device, so it costs us nothing per song and we can leave it free. There is no sign-up, no credit counter and no daily cap. The limits are your own device and the length of a song: 10 minutes in Fast and 6 in HQ. Your song is never uploaded. What we do download, once, is the code that runs the models and the model you choose (85 MB for Fast, 353 MB for HQ), from public hosts.

It is a beta: it works, but some devices will be slow. If it breaks on your song, write to info@obra.art.