How to Make a Karaoke Track: Vocal Removal, Bleed, and Rights | WaveCraft

The short answer

Run vocal separation on the song, keep the instrumental.wav it produces, and that file is the karaoke backing. One pass, run entirely on your iPhone, so the audio never leaves the device. The mechanical part is easy. The two things people get wrong are that a little of the original vocal survives — separation reconstructs the accompaniment rather than extracting it — and that removing a vocal from a song you do not own does not make the result yours to publish, however cleanly it comes out.

WaveCraft is on the App Store if you want to skip ahead. The rest of this page is about what makes a karaoke track actually usable, and where the legal line sits.

Why you are actually here

Three underlying problems hide behind this search, and they need different answers:

• A performance is coming up and no karaoke version exists. You need a backing in your key, at a level you can sing over.
• You want to practise a part. You do not want a backing — you want to hear what the original singer did.
• You make karaoke or cover content. Here the technical question is nearly irrelevant; the rights question decides whether the video survives its first week.

A karaoke track is not a "vocal removed" file. It is a finished track with a specific job: a different singer has to perform over it — and that is stricter than it sounds.

Why the original singer is still in there

Separation estimates rather than extracts. WaveCraft converts the mix to a spectrogram with a 4096-point FFT, and asks a neural network — Open-Unmix, running on-device via Core ML — to estimate the vocal's magnitude at every point in time and frequency. A soft mask splits that estimate into the vocals file and its complement, the instrumental.

The mask is a per-pixel judgement, not a subtraction. Where the model is confident the vocal dominates, the vocals file takes nearly all the energy. Where the vocal and the guitars share a band at the same moment, it has to guess, and guesses are visible:

• Sustained notes. A held vowel is broadband and sits on top of the chord. This is where bleed survives.
• Reverb tails. The vocal's reverb belongs to the room, not the voice, so the mask cannot assign it cleanly. The last word of a line leaves a ghost.
• Dense, loud mixes. Compressed masters push the vocal into the instruments rather than in front of them — the case with the least evidence to work from.

Two consequences follow from the technique, not from a bug:

The two files add back up to the original. Every mask value sits between 0 and 1 and the pair sums to exactly 1; WaveCraft's tests assert this to within 0.001 relative error. Stacking both stems reproduces your source sample for sample. That is also why the instrumental cannot be cleaner than a perfect removal — it is the original minus an estimate.

The karaoke track will be quieter than the record. A lead vocal typically carries a large share of a mix's perceived loudness, so an instrumental-only mix lands well below where the original sits. This is the most common complaint about home-made karaoke tracks, and it is not a defect.

Making it singable, not just separated

Once the instrumental exists, the rest is ordinary editing — and this is where a phone editor earns its place:

• Trim what you can hear. A fade on the last two seconds takes the outro reverb ghost with it; a short fade at the head removes any first-word artifact.
• Cut a frequency, not the level. Residual voice usually sits in the 1–4 kHz band where vocal intelligibility lives. An EQ cut there removes more than any volume change and leaves bass and cymbals alone.
• Pull the whole track down 1–2 dB. A karaoke bed is a support, not a competitor. Highest impact per drag of any control here.
• Normalize loudness to a target in LUFS, so the bed sits correctly whether you started from a quiet indie track or a brick-limited pop single.
• Fix the key instead of giving up. Pitch shift retunes the instrumental without changing tempo.

WaveCraft's twelve effects span EQ, dynamics, space, modulation, drive and pitch/time, plus mastering presets. A karaoke track rarely needs more than an EQ, a limiter and a volume setting — the point is that all three live in the same project as the track.

Where this fails, and the honest ceiling

A 48 kHz source is normalized down to 44.1 kHz. The vocal model was trained at one sample rate and one FFT geometry; changing either feeds the network input it has never seen. Nobody hears this under a song, but a 48 kHz session needs a resample on export.

Mono sources still work. Karaoke from a mono recording is common — old demos, voice notes, rehearsal-room takes. The model expects two channels, so WaveCraft duplicates the mono signal into both and produces mono stems rather than failing.

Heavy distortion and screaming separate worst. Belting, harsh consonants and saturated vocal processing put energy exactly where the instruments are.

And the one that catches people out: a separated instrumental is not a note-perfect backing. It is the original mix with the center removed, so where the singer was loud and bright the bed feels thin — instruments competing for that space were partly assigned to the vocals file. Commercial karaoke beds are played by session musicians, not produced by separation.

The rights question, which is the real one

Personal, non-commercial karaoke — singing at home, practising, a private party — is generally treated as fair use in most jurisdictions, including the US.

Anything public changes the analysis. Separating the vocal does not change the copyright status of the underlying recording. You still hold a copy of someone else's copyrighted music, and platforms run fingerprinting that matches the recording, not the vocal melody. A technically perfect instrumental still gets claimed, because the claim is not about whether a voice is present.

The familiar workarounds do not hold up:

• "I removed the vocals, so it's a new work." No. The recording is the protected work and is still recognisable inside your file.
• "I slowed it down or pitch-shifted it." Arguable in some cases; still a claimable recording.
• "It's only 30 seconds." Duration is one factor among many, not an exemption.
• "It's a cover." More complicated, not simpler: composition and master recording belong to different rights holders, and a cover clears neither.

What works: your own compositions, or material where you hold both composition and recording; royalty-free libraries that explicitly grant the use — many grant personal use only, so read the licence rather than the category name; official instrumentals or stems published by the rights holder, which some artists and labels release precisely to enable karaoke and covers; and public-domain compositions, with the caveat that a modern recording of one is separately copyrighted — the composition is free, that performance usually is not.

If you publish karaoke as a business, licensing is cheaper than disputes: one takedown costs more than a year of licence fees, and a claim on a monetization-linked account reaches beyond the offending video.

Practising: the other output is more useful

If the goal is learning a part rather than publishing, the vocals file is worth more than the instrumental. Slow it with time stretch, transpose it with pitch shift, loop the four bars you keep fumbling, and practise against the actual performance — real timing, real breath, real resolution — rather than your own imitation. That is why the complementary split matters: you keep both halves.

What WaveCraft does differently

WaveCraft is a multitrack audio editor whose separation runs on the device rather than a server:

• The audio never leaves the phone. No upload, no queue, no account, and it works with no network at all. Separating a song you are about to post publicly is a decision about where a copyrighted file goes; here the answer is nowhere.
• Separation is one tap inside a real editor, not two files you then hunt for software to clean up. The stems land as tracks, where the EQ, fades, loudness normalization, limiting and pitch shift are one tap away.
• The mask is complementary, which is what makes the practice workflow above real rather than theoretical.
• Lossless export, no surprise formats. WAV, AAC, ALAC or FLAC, up to 32-bit float, with 16 and 24-bit available. There is no MP3, because iOS has no native MP3 encoder — you get formats that are actually encodable rather than a lossy one chosen for you.
• Free where this workflow needs it. Multitrack recording and editing, every effect and preset, and WAV/AAC/ALAC/FLAC export are free and stay free. Vocal separation, auto-mix, the mastering chain, stem export and batch export are WaveCraft Pro, available monthly, yearly or as a one-time lifetime purchase with Family Sharing.

The honest limit: this is a vocal split, not a full stem splitter — you get vocals and everything else. Separating drums, bass and harmony individually is a different and much heavier operation on a phone, and for karaoke a two-way split is exactly what the job requires.

What to decide now

• Practising or private singing — this works today. Separate, keep the instrumental, trim the tail, sing. Keep the vocals file too; slowing it down teaches you more than the backing will.
• Posting publicly — the risk is real and it attaches to your account. Use your own material, a licence that grants it, or an official instrumental.
• A one-off party or class — a personal-use licence from a royalty-free library, or a song whose label ships an official karaoke version.

Not worth doing is separating the track and sorting out the rights afterwards. That inversion is the only mistake here that can cost you something.