How to Make a Karaoke or Instrumental Version of a Song with StemRoller

Updated

StemRoller creates a karaoke-ready instrumental automatically: type a song name into its search bar, and the app finds the track online and splits it into four stems — vocals, drums, bass, and everything else. It then mixes all the non-vocal stems into a single instrumental track, so you get a backing track without any manual setup. This works for any song StemRoller can locate online, and the app is free and open-source, built on Facebook's Demucs AI research project.

What you get from one search

A single search produces five audio files:

Track Contents Typical use
Vocals Isolated singing/voice Acapella, remixing, practice
Drums Isolated percussion Sampling, drum replacement
Bass Isolated low-end Sampling, transcription
Other Everything else (guitars, keys, synths, etc.) Sampling, remixing
Instrumental All non-vocal stems mixed into one track Karaoke, singing or playing over

The instrumental is the file you want for karaoke — it is generated for you, not something you have to assemble yourself.

Steps

  1. Open StemRoller and find the search bar at the top of the window.
  2. Type the name of the song you want. StemRoller searches for it online rather than requiring you to supply an audio file.
  3. Start the split. The app runs the Demucs model and separates the recording into the four stems listed above.
  4. Wait for processing to finish. Separation is AI-based and takes some time per song; longer tracks take longer.
  5. Locate the output files. The instrumental track is the one to use for karaoke; the vocal track is useful if you want an acapella or to check the separation quality.

Points to check before you rely on it

  • Source quality matters. StemRoller pulls the song from online sources, so the separation quality depends on what it finds. A clean studio master separates better than a live recording or a heavily compressed upload.
  • Separation is not perfect. AI stem splitting can leave artifacts — bleed of vocals into the instrumental, or instruments leaking into the vocal track. For casual karaoke this is usually fine; for release-quality work, expect to clean up the results.
  • No manual configuration is needed. There is no stem-count or model-setting step described; the four-stem split plus instrumental is the default output.
  • Support and licensing. The project is open-source and free, with donations accepted. If you run into problems or have questions, the project points users to its Discord community. Check the licensing terms of any song before publishing or performing a derived track.

If something goes wrong

  • Song not found: try a more specific or differently formatted title, since the app depends on locating the track online.
  • Poor separation: try a different source version of the same song, or accept that some tracks (dense mixes, live recordings) separate less cleanly.
  • Questions or bugs: the project directs users to its Discord for help.