Published: 2026-06-10
Karaoke night, a cover video, band rehearsal without a singer — sooner or later everyone needs a version of a song without the lead vocal. A few years ago that meant hunting for an official instrumental or fighting with EQ tricks that ruined the mix. Today an AI model does it in minutes, from any recording you have. Here is the whole process, step by step.
The quality of the karaoke track is capped by the quality of the original. Use the best file you can find: a lossless wav or flac is ideal, a high-bitrate mp3 (256–320 kbps) is perfectly fine. Avoid videos re-recorded from a phone or low-bitrate rips — separation models amplify whatever artifacts are already in the file.
Open the TrackStemLab karaoke maker, sign in and upload your file on the Separate page. Under the hood the service runs Demucs — a state-of-the-art neural network that has learned what a human voice sounds like inside a mix and pulls it out, leaving drums, bass and instruments untouched. Processing takes a few minutes for a typical song.
If you only need vocals gone, a 2-stem model (vocals + instrumental) is the fastest option. Multi-stem models also give you drums, bass and other instruments separately — useful if you later want to adjust the balance of the backing track.
Listen to the instrumental from start to finish in the browser preview. Two things to check: leftover vocal traces in the choruses (dense harmonies are the hardest case) and the overall tonal balance. Modern models keep the original stereo image, so the result should sound like the official instrumental, not like a hollowed-out radio edit.
The separation gives you the isolated vocal as a bonus. Keep it: it works as a guide track when rehearsing, helps you learn the exact phrasing, and if you produce a cover you can study how the original singer sits against the beat.
Classic vocal removers invert one stereo channel or cut mid frequencies. They only work when the voice is mixed dead-center and untouched by stereo effects — which is almost never true on modern records — and they always damage the instruments that share the center: kick, snare, bass. AI source separation has none of those constraints; we cover the details in Demucs vs classic vocal removers.
The separated track is still the original recording. Singing over it at home or at a private party is one thing; publishing a cover or performing publicly may require licenses depending on your country. When in doubt, check the rules that apply to you.
Ready to try? Create a free account — every account gets free processing minutes each day, enough to make karaoke versions of a few songs.
← All articles