AI Stem Splitter
Separate vocals, drums, bass, guitar and more. Upload your audio, listen to each part, and download lossless WAV stems. Completely free. No account needed.
Drop your audio here
WAV, MP3, FLAC, M4A, OGG, AIFF, AAC ยท Up to 10 minutes / 100 MB
Upload only audio you have the rights or permission to process.
Files stay private. Download within 24 hours. Files are automatically deleted afterwards; they are not used for training.
Eight parts. One upload.
How AI stem separation works
Source separation estimates individual vocal and instrument tracks, called stems, from a mixed recording. This tool uses X-LANCE models to extract eight instrument groups. At their core is BS-RoFormer, a neural network that models relationships across frequency bands and time.
1. Represent sound in time and frequency
A short-time Fourier transform (STFT) analyzes brief windows of audio to describe the frequencies present at each moment. It includes both amplitude, the strength of the sound, and phase, the position of its waveform cycle. This representation is split into frequency bands and converted into features the network can process.
2. Connect patterns across time and frequency bands
BS-RoFormer uses attention to relate information along both axes. For overlapping sounds such as bass and kick drum, attacks, sustain, harmonics and changes over time provide cues. The network estimates a complex mask that modifies amplitude and phase, then an inverse STFT converts the result back into audio.
3. Extract each instrument group with dedicated models
Each of the eight groups starts from the original mix. For drums, synth, percussion and orchestra, the output of a first model feeds a second model for further extraction. Vocals, bass, guitar and keyboard each use a model dedicated to that group.
4. Join overlapping segments and export editable WAVs
Long recordings are processed in 10-second segments with a 2-second overlap. Gradually changing weights blend adjacent segments to reduce discontinuities. Outputs are stereo, 44.1 kHz, 32-bit float WAVs with matching start positions and sample counts. No per-stem loudness normalization is applied, so you can place them at the same point on a DAW timeline for editing.
Using your separated audio
Upload an audio file, let AI estimate each instrument group, then listen or export to your DAW. Useful for studying arrangements, learning parts and working with recordings you own.
Separation quality and use
Instruments sounding at the same time in similar frequency ranges, heavy distortion and diffuse reverb make separation harder. You may hear leakage from other instruments, missing details or smeared sound. Categories such as keyboard and synth can overlap. Because each group is estimated independently, summing all eight stems may not exactly reproduce the original mix. Listen both in solo and in the combination you intend to use.
Separation does not grant new rights to the recording or composition. Use results only as permitted by the relevant rights holders. This tool does not create a new performance or clear music rights.
File storage and deletion
Audio is processed on Modal and temporarily stored in private Google Cloud Storage. Results are available to your browser for 24 hours. Automatic deletion follows; you can also delete completed files immediately. We do not use audio for training. A browser cookie and a hashed network identifier protect jobs and prevent overlapping requests.