StemSplit
Make a karaoke track by removing the vocals, pull the vocals out on their own, or split a song into stems and remix them, then loop, slow down or change the key to practise. An optional on-device AI mode separates vocals, drums, bass and the rest; it works best on studio stereo mixes.
Runs in this tab. Nothing is uploaded. Keeps working if your connection drops.
Drop a song here, or choose one
MP3, WAV, M4A, FLAC, OGG and anything else your browser plays. It never leaves your device.
Or try a sample
Each sample is a short stereo song made in your browser, panned the way studio mixes are.
- Karaoke or a cappellaTurn the vocals off, or hear only them, in one tap.
- A four-stem mixerVocals, instruments, bass and highs, each with mute, solo and volume, live while it plays.
- Optional AI stemsDownload the AI models once (85.5 MB) for real vocals, drums, bass and everything else, split on your device.
- Loop, slow down, change keyLoop a section, slow it down to practise, or move it into your key without changing the speed.
- See what changedA spectrogram compares the original with your mix, or shows only what was taken out.
Works best on studio stereo mixes with the vocal in the centre. This is classic signal processing (stereo-centre extraction and crossover filters), not AI separation. Anything else in the centre of the voice band, such as a snare or a centred piano, comes out with the vocal, and reverb on the voice stays behind with the instruments. Mono files can't be split into vocals and instruments.
AI stems are optional. They use Open-Unmix (MIT licence), trained on pop and rock recordings, and handle reverb and mono far better, but expect some bleed between stems. The models are a one-time 85.5 MB download from this site; a 3-minute song takes about a minute on a laptop (longer on phones) and needs about 900 MB more memory.
The song is split on your device at its own sample rate, and exports keep that rate. It needs about 104 MB of memory per minute of 44.1 kHz stereo audio (186 MB with key or tempo changed); songs longer than this device can likely handle get a warning first. Exports are 16-bit, 24-bit or 32-bit float WAV files. Songs must be at least 0.1 s long.
