Stella
Audio and music

Remove the vocals from a song for karaoke

A song with the singing removed and the bass and drums kept, for practice.

Download StellaFree with the plan you already have.

A real run on 13 November 2026, on a Linux computer, using Claude Code with Opus, on a Claude plan. The waiting is sped up; it took 3 min 19 s in real time.

The short answer.

Use an AI separation model, not the old centre-cancel trick. The free Ultimate Vocal Remover app runs models like Demucs on your computer; sites like LALAL.AI and Moises do it online with limits. Audacity's Vocal Reduction also wipes out bass and drums. Stella (stella.sh) installed the open-source Demucs model and removed the vocal from a song while keeping its bass and kick, then checked the result for leftover voice.

Stella does this on Windows, Mac and Linux.

What happened.

The karaoke version in about 3 minutes 20 seconds, including the install. Checked afterwards against an answer key written before the run, using the song's true instrumental: the result is 15 dB closer to the instrumental than the original was (6 dB required), the bass and kick are fully kept (not the centre-cancel trick), and the original is unchanged.

The backing track was synthesised for this run and the vocal was sung by a speech model, so the true instrumental was known for checking. Stella also noticed a second copy of the song elsewhere on the test computer and asked about it; that folder path is blurred in the recording. The work is sped up 13×.

How Stella does it.

  1. Finds “Lantern Song (demo)” in Music
  2. Sets up the open-source Demucs separation model in its own Python environment, a download of several gigabytes including PyTorch and the model
  3. Splits the song into vocals and everything else with Demucs's fine-tuned model, then runs the backing track through again to check for leftover voice
  4. Saves Lantern Song (demo) - Karaoke (no vocals).mp3 next to the original, at the original volume, and leaves the original unchanged
  5. Says the result was measured, not listened to, and that parts near the singer's pitch may sound a little thin

The prompt.

Make a karaoke version of Lantern Song in my Music folder with the singing taken out, so I can practise singing over it.

Paste it into Stella and press Enter.

This is the exact prompt from the recording. If your files are somewhere else, say where.

Do it yourself.

  1. Ultimate Vocal Remover (Windows, Mac, Linux)

    Download the free Ultimate Vocal Remover from its GitHub releases page. Choose the song as input, pick a Demucs or MDX-Net model, tick Instrumental Only, choose an output folder and press Start. The first run downloads the model.

  2. Demucs from the command line

    With Python installed, pip install demucs, then demucs --two-stems=vocals "song.mp3" writes vocals.wav and no_vocals.wav into a separated folder. It runs on a normal processor, faster with a graphics card.

  3. Audacity: free AI plugin, or the quick trick

    Audacity's free OpenVINO plugins add a Music Separation effect on Windows and Linux. Its built-in Vocal Reduction and Isolation cancels whatever is in the centre of the stereo mix, which removes the bass and kick drum along with the voice.

  4. Online vocal removers

    LALAL.AI, Moises, vocalremover.org and others separate in the browser. They're quick but upload the song and limit free minutes or quality.

  5. Listen for leftovers

    Play the result through. Faint echoes of the voice, reverb tails, or a thin sound where an instrument shares the singer's range are normal side effects. Trying a different model sometimes helps.

Where it stops.

  • Separation is very good but not perfect: faint vocal echoes and reverb can remain, and instruments close to the voice may sound thinner. The check measured the result; nobody listened to it in the recording.
  • It's a big install: Stella's private Python environment for Demucs took about 5 GB, mostly PyTorch, plus a 320 MB model. Ask Stella to delete it afterwards if you won't need it again.
  • For personal practice. Sharing a karaoke version of someone else's song is a copyright question, not a technical one.
  • The song's name goes to the AI model you use in Stella; the song itself is separated on your computer and isn't uploaded. Recorded on Linux with a graphics card; on a Mac or Windows PC the same model runs, more slowly without a supported graphics card, which wasn't recorded.

Questions.

Can you really remove vocals from a song?

Mostly. AI models like Demucs separate the voice well; in the test the backing track came out 15 dB closer to the true instrumental than the original. Some faint traces can remain.

Why does Audacity's vocal remover kill the bass?

It subtracts the left channel from the right, removing anything mixed in the centre. Vocals are centred, but so are bass and kick drum.

Can I keep just the vocals instead?

Yes. The same separation produces a vocals-only track; ask Stella to save that too.

Can it also change the key to suit my voice?
Is Stella free for this?

Yes. Free with the ChatGPT or Claude plan you already have. Sign in with your ChatGPT plan, or use your Claude plan through Claude Code, and this task costs nothing extra. New accounts also get a small free trial of Stella’s own models.

Can I control what Stella changes on my computer?

Yes. Stella works on Windows, Mac and Linux with your own files and shows each step in the chat, so you can watch it and stop it at any point. Tell it what it may not do, as the prompts on these pages do (“don't delete anything without asking”). In the Downloads recording it deleted nothing until it was told to go ahead. It also refuses commands that would wipe a whole disk, your home folder or system folders.

Checked against these sources on 13 November 2026.

Stella