Split a song into stems
Upload a mix and get isolated tracks back — vocals, drums, bass and the melodic parts, each as its own file. What makes this different from a standalone splitter is where the stems go next: straight into a real session, on the grid, with the tempo and key already worked out.
What you get back
A separated mix comes back as individual audio files, one per source. Vocals and drums are the cleanest — they are the most distinct in a mix and the model has the most to work with. Bass is reliable. Guitars and keys are harder, because in a dense arrangement they overlap in both frequency and time, and what you get is a best effort rather than a surgical extraction.
Every stem keeps the timing of the original. Nothing is time-stretched or re-quantised on the way out, so the files line up sample-accurate against the source when you drop them on a timeline.
Why the separator is not demucs
A lot of tools in this space wrap the same open-source model. This one runs a RoFormer-based separator, which is a transformer operating on the spectrogram rather than a convolutional network on the waveform. In practice the difference shows up on the hard material: dense mixes, heavy compression, and vocals that sit inside a wall of guitars.
It is not magic. Separation is a fundamentally ill-posed problem — once two instruments have been summed into one channel, the information to pull them apart perfectly is gone. Anyone claiming otherwise is selling something.
The part that usually costs you an afternoon
Most separators hand you a folder of WAVs and stop. That is the point where the actual work starts: importing them, finding the tempo, lining up the grid, naming tracks.
Because separation here is a stage in a pipeline rather than the product, the same upload can come back as a finished session file instead — tracks laid out, tempo detected, each stem on its own named track. That is the audio-to-session route, and it is the one most people actually want.
FAQ
How many stems do I get?
Vocals, drums, bass and the remaining melodic material, with further splits available depending on the source. A sparse arrangement separates further and more cleanly than a dense one.
Does it work on a live recording?
Yes, though bleed between sources makes it harder — a room recording where every mic hears every instrument is the worst case for any separator, including this one.
Can I get the stems as MIDI instead of audio?
Yes. Transcription runs per-stem, so you can take the separated parts through to notes rather than audio. That is a different route through the same pipeline.