How to Remove Vocals from a Song: Karaoke Tracks, A Cappellas and Stems
AI source separation such as Demucs splits a finished mix into vocals and instrumental, or drums, bass and other stems. Start from the best-quality file, listen through the loudest chorus for traces, and use it only on recordings you have the rights to.
Fastest way
Ask your AI to do it
Paste this into the AI agent you already use:
Help me use a Skill with another agent. Ask which agents and package I use, check supported locations and tools, identify any assumptions to change, and verify the workflow with a small example.
Quick answer
The old trick of cancelling whatever sits in the centre of a stereo mix also removes the bass and kick and leaves reverb behind. Modern separation models are trained on thousands of multitrack songs and estimate each part directly, so the instrumental keeps its low end and the vocal comes out on its own.
Get the cleanest result
- Use WAV or FLAC rather than MP3; compression artefacts end up in both parts.
- Make the instrumental by subtracting the vocal, so the two always add back up to the song.
- If traces remain, try a fine-tuned model or average several passes.
- Listen on headphones through the loudest chorus before you use it.
Practice and karaoke versions
A practice track keeps a quiet guide vocal at about a quarter of its level. A karaoke version in another key shifts the pitch while keeping the tempo; two semitones either way suits most voices without obvious artefacts.
Rights
Separated parts of a song are still that song. Use them for your own recordings, licensed tracks, private practice, or work you have cleared with the rights holders.
Do it with an AI agent
VocalSplit runs Demucs on your own computer to make an instrumental, an a cappella and stems, with practice mixes, key changes, loudness checks and an offline mixer to compare every part.