How to remove vocals from a song: steps and limits

To remove vocals from a song you need a model that separates the voice from everything else, not a filter. Our vocal remover does that in the browser: upload a track, and it returns six stems: vocals, drums, bass, guitar, piano and other. Mute the vocals and what is left is the instrumental. Below is how to use it, what you will get back, and where it falls short.

How to remove vocals from a song, step by step

1. Open the vocal remover and drop in your file. It accepts MP3, WAV, FLAC, M4A, OGG, AAC and AIFF, up to 10 minutes and 120 MB.
2. You get a free 60-second preview with no account. Listen to it first: if the instrumental is not good enough on this song, you have not spent anything.
3. If it is good, sign in to split the whole song. Nothing about the file needs to change.
4. Mute the vocals in the mixer to hear your song with the vocals taken out, and download the other stems as 320 kbps MP3s. The vocals stem is the voice on its own.

The results of a full split also show the song's tempo and key. If you only want those, the BPM finder and key finder are free and need no signup.

Why the old phase-cancellation trick rarely works

You may have seen the classic advice: invert one channel of a stereo file and mix it with the other. This cancels whatever is identical in both channels. Lead vocals are often panned to the centre, so they cancel, but so do the kick, the bass and often the snare, and any vocal with stereo reverb leaves a ghost behind. A separation model instead learns what a voice looks like in the spectrogram and removes that, so the drums and bass stay.

How good is the instrumental?

We measured the vocals-and-instrumental model StemX used until October 2026 (UVR MDX-NET) on 12 tracks from the MUSDB18 test set, using scale-invariant SDR (higher is better; what SI-SDR means). Medians:

The instrumental scores 14.15 dB, so removing the voice is the stronger half of the job. It is not the top instrumental score we measured: other models in the same family reached 14.97 to 15.21 dB on the instrumental but scored only 6.00 to 6.38 dB on vocals, so we shipped the one that was better overall. The full comparison is in our benchmark post, and why StemX offered only vocals and instrumental until then is in why we shipped two stems and not six.

Limits you should expect

• Artefacts. Separation is not subtraction. On dense mixes, with heavy reverb, harmonies or a voice buried in loud guitars, you can hear faint remnants of the singer or a watery quality in the backing. A good score is not a perfect one.
• Small test. Twelve seven-second excerpts is a small benchmark. Read the numbers as a guide, not a guarantee for your song.
• Six stems, no more. You get vocals, drums, bass, guitar, piano and other. Anything else, such as strings or synths, lands in other.
• Source quality. The result cannot be better than the file you give it. A low-bitrate upload gives a lower-quality instrumental.

Use tracks you have the rights to

Only upload music you own or have the rights to use, especially if you plan to publish the result. Removing the vocals from a song does not change who owns the recording or the composition.

Credit

Separation uses BS-RoFormer-SW (bs-roformer-infer, MIT; weights enerjazzer/BS-ROFO-SW-Fixed); HT-Demucs 6s (Meta AI Research, demucs MIT). The lead and backing vocal split uses UVR models by Anjok07 & aufr33, released under the MIT licence with credit to UVR and its developers.

Ready to try it? Remove the vocals from your song.