Split a Song into Stems

Break a finished song back into its parts for remixing, re-balancing or study. Separation runs on open-source models with commercially licensed weights: Mel-Band RoFormer for vocals and SCNet for drums, bass and other.

AI model This tool runs a generative AI model on a GPU, so it takes longer than the instant filters. Free to try; paid plans get priority on the GPU.

🎧

Drop an audio or video file here

or

MP3, WAV, M4A, FLAC, OGG, AAC

Cleaning your audio…

Your result link
Before

Tip: press the space bar to toggle Before / After.

Download stems
Download cleaned audio Sign up for full-length files

How it works

Mel-Band RoFormer separates the vocal from the instrumental, then SCNet splits that instrumental into drums, bass and other. You download four files: vocals, drums, bass and other, which add back up to the original mix. Each model learns what its sources look like in a finished mix, so it works on any song rather than relying on panning tricks.

What it's good for

  • Remixing and production
  • Music education and analysis
  • Re-balancing a stereo mix
  • Sampling individual instruments

Details

Engine
Mel-Band RoFormer + SCNet
Formats
MP3, WAV, M4A, FLAC, OGG, AAC
Price
Free to try

Frequently asked questions

Four separate files: vocals, drums, bass and other (everything that is not vocals, drums or bass, such as guitars, keys and synths). Played together they add back up to the original mix. For a single instrumental file, use the vocal remover tool.

On 50 excerpts from the MUSDB18 test set we measured a vocal SDR of 11.3 dB (higher is cleaner), against 8.4 dB for the model we used before. Densely layered parts can have slight bleed.

Yes, roughly twice as long. It runs the same vocal separation pass as the karaoke tool and then a second model that splits the instrumental into drums, bass and other.

Yes. Every split gives you four downloads: vocals, drums, bass and other, each as its own file in the format you chose.

About twice as long as a vocal isolation, because a second model splits the instrumental into drums, bass and other after the vocal pass; longer tracks take proportionally longer.

Stereo input helps the model place instruments and gives cleaner stems; mono files still separate but with more cross-bleed between parts.

Technically you can remix them, but the source recording is copyrighted, so a released remix or sample still requires clearance from the rights holder.

Related tools