Stems Separator
Pull the voice out of a finished mix. Upload a track and get two files back — the isolated vocal and the instrumental it was sitting on.
Separation costs 2 credits. Failed jobs are refunded automatically.
What is stem separation?
Stem separation takes a mixed-down recording and estimates what the individual parts sounded like before they were summed together. A finished master is one stereo file; separation reconstructs the vocal on its own and everything else on its own, so you can work with a part that was never delivered to you as a separate track.
This is estimation, not un-mixing. There is no hidden copy of the vocal inside an MP3 to extract — a neural model trained on paired mixes and stems predicts what each source contributed to every moment of the audio, and writes that prediction out as new audio. Sparse arrangements with a clear lead vocal come back cleanest; dense, heavily-compressed masters leave more traces of one stem in the other.
This tool separates two stems: vocals and instrumental. The instrumental is the model's drums, bass and other outputs summed back together, which is the same thing you get from a two-stem run of the underlying model — no separate drum or bass files.
Stems separator features
Two stems, vocals and instrumental
One upload returns exactly two files: the isolated vocal and the backing track. No stem picker, no four-way split to sift through.
Demucs htdemucs model
Separation runs on the hybrid transformer Demucs model on a GPU — the same open model that sets the reference quality bar for source separation.
Six input formats
MP3, WAV, FLAC, AIFF, OGG and M4A are all accepted. Files up to 25 MB and 6 minutes long.
MP3 downloads
Both stems come back as 192 kbps MP3s you can play in the browser and download straight away.
Nothing to install
The separation runs on our GPUs, not your laptop. No Python environment, no model weights to download, no plugin to license.
Included with Producer
Stem separation is a Producer-plan feature and costs 2 credits per track. Failed jobs refund automatically.
How stem separation works
1. Upload a mixed track
MP3, WAV, FLAC, AIFF, OGG or M4A, up to 25 MB and 6 minutes. The file is decoded to raw audio before anything else happens.
2. The model separates the sources
Demucs estimates the vocal, drums, bass and remaining instruments across the whole track. Everything that is not the vocal is summed back into one instrumental.
3. Play and download both stems
The vocal and the instrumental appear as two players on the page, each with its own MP3 download.
Stems separator FAQ
How do I separate vocals from a song?
Upload the track on this page and the separation runs automatically, returning a vocal stem and an instrumental stem. Nothing is installed locally — the model runs on our GPUs and hands you two MP3s.
How clean is the result?
Clean enough to sing over, sample or study, but not identical to the original multitrack. Reverb tails on a vocal tend to follow the vocal into its stem, and loud cymbals or synth pads sometimes leave a trace in the vocal. Sparse mixes with a clear lead separate best.
Can I get separate drum and bass stems?
Not here. This tool is deliberately two-stem: vocals and instrumental. The underlying model does estimate drums and bass separately, but we sum them into one backing track because that is what the overwhelming majority of separation jobs actually need.
What are the file limits?
25 MB and 6 minutes. Longer tracks are rejected rather than truncated, so you always get either the whole song or a clear error — never a half-separated file you did not ask for.
Does separation cost credits?
Yes — 2 credits per track, and stem separation is included with the Producer plan. If a job fails, the credits are refunded to your balance automatically.
Is it legal to separate a song I did not make?
Separating a track does not give you rights to it. Using stems from someone else's recording in a release, or distributing them, needs permission from the rights holders in the same way sampling does. Practice, study, karaoke and personal use are a different matter, but the copyright in the original recording does not disappear because the parts were split.
Split a track into vocals and instrumental.
One upload in — two stems out, ready to play and download.
More AI music tools
One place to generate, convert, detect, and remix.
Turn a prompt into a full song with vocals.
Write full song lyrics from a theme or a line.
AI rap diss with lyrics + beat in seconds.
Stage name ideas by scene — trap, drill, classic.
Convert an MP3 into editable MIDI notes.
Add a slowed + reverb tape vibe, free in your browser.
Find any track's tempo by tap or upload.
Find whether a song counts in 4 or in 3.
Detect the musical key of any song.
Check your pitch accuracy note by note.
Get key, Camelot code and tempo from one upload.
Batch-key up to 25 tracks into one table.
Find the key through your mic — no file needed.
Get every beat position, bar by bar.
Tap out a tempo, get delay times in ms.
Every note and chord in any key.
Find the note and frequency of any sample.
Live chromatic tuner in your browser.
Turn a melody into note names and timings.
Measure your vocal range low to high.
WAV, FLAC, AIFF, OGG, M4A or MP3 into editable MIDI notes.
Piano audio to engraved sheet music, both hands.
Any recording into a readable, printable score.
AI vs by ear vs hiring — and the fast path.
6 formats in — score, PDF, MusicXML, MIDI out.
Tape warmth, wobble and vinyl crackle on any track.
Guitar, bass, ukulele and more — notes, Hz and a tuner.
BPM to ms for every note value.