Why Your Vocal Remover Leaves Artifacts

The watery, swirly, "underwater" sound in your separated track is residue from an educated guess. When a free vocal remover can't cleanly decide whether a given sliver of sound belongs to the voice or to the backing, it splits the difference — and that half-in, half-out energy is exactly what you're hearing. It's not a broken file, a bad export, or a setting you missed. It's a side effect of how AI separation works, and while you can't erase it entirely, you can meaningfully reduce it.

This guide is about that specific problem: the artifacts, where they come from, and what actually helps. If you want the full picture of the mechanism first, how AI vocal removal actually works is the companion explainer — this one assumes you've already got a result in front of you that sounds a little off.

What you're actually hearing

The telltale signs cluster into a few textures, and they usually appear together:

• A watery or phasey swirl that rides on top of sustained notes and held chords.
• An "underwater" muffling, as if the audio is playing through a thin membrane.
• Faint pumping or breathing at a steady rhythm, most noticeable in quiet passages.
• Vocal ghosts — wisps of the singer left behind in an instrumental, or a smear of backing left in an acapella.

Each of these traces back to a different stage of the separation. Knowing which is which tells you what's worth trying and what's a dead end.

Why a vocal remover leaves artifacts

The core reason: overlap in the same time-frequency bins

An AI vocal remover doesn't listen to your song the way you do. Part of how these models work is to analyze the audio as a spectrogram — a picture of which frequencies are loud at each moment in time — and decide, for every tiny time-and-frequency tile, how much of it belongs to the vocal versus everything else. Sort the tiles one way and you rebuild the voice; sort them the other way and you rebuild the instrumental.

That works beautifully when a sound sits in its own patch of the spectrogram. The problem is that music constantly stacks different instruments into the same tiles. A sibilant "s" and a crash cymbal live in the same high-frequency region. A held vocal note and a sustained synth pad occupy the same band for seconds at a time. When two sources share one tile, there is no correct way to divide that energy — the model has to guess a split, and an imperfect guess leaves a bit of the voice in the instrumental and a bit of the pad in the acapella. Multiply that across thousands of contested tiles and you get the shimmering, watery residue.

This is the part worth internalizing: the artifact isn't a bug the model will one day patch away. It's inherent. Wherever two sounds genuinely overlap, some ambiguity is unavoidable, and better models mostly just shrink it rather than remove it.

Phase: why it sounds "underwater" specifically

There's a second reason, and it helps explain the underwater quality in particular. When a model works on the spectrogram, it's mainly deciding how loud each frequency should be — its magnitude — but a spectrogram also carries phase, the timing alignment that tells the waveform how to reassemble. If the magnitudes are reshaped while the phase stays closer to the original mix, the reconstructed audio smears slightly in time. That smear is a big part of the muffled, submerged texture you hear on separated parts.

Chunk boundaries: the faint pumping

Separation runs on a long track in overlapping windows rather than all at once, then stitches the pieces back together (an approach called overlap-add). The seams between windows aren't always perfectly matched, and a small mismatch can surface as faint periodic pumping or swirl at a steady interval. It's usually subtle, and it's often what people mistake for a rhythmic "breathing" in an otherwise clean stem.

The material itself makes it better or worse

Two songs run through the same tool can come out wildly different, and it's mostly the input's fault, not the model's:

• Lossy sources start behind. A low-bitrate MP3 has already thrown away high-frequency detail and baked in its own compression artifacts. The separator can't recover what the encoder deleted, and it can amplify the encoder's damage while trying. Feed it that and the water gets louder.
• Effects blur the boundaries. Heavy reverb, doubled takes, and wide stereo processing spread a voice across time and space, so its "fingerprint" overlaps more of everything else. (This is the same trait that broke the old phase-cancellation trick — a doubled, reverb-drenched vocal was never cleanly centered to begin with.)
• Dense masters overlap more. A sparse arrangement leaves each instrument room in the spectrogram. A modern wall-of-sound master, limited to be loud with everything stacked, packs far more sources into the same tiles — more contested territory, more residue.

That's why a stripped-back acoustic track can separate almost cleanly while a maximalist pop single sounds watery no matter what you do.

How to reduce the artifacts

Separation is never perfect, so the honest goal is less bleed, not none. In rough order of impact:

Start from the best source file you have. This is the single biggest lever. A lossless WAV or FLAC — or at least a high-bitrate MP3 — gives the model detail a low-bitrate file simply doesn't contain. Everything here acts on a file you already have, so if you have access to a better-quality copy of the same song, use that one.

If you only need the instrumental, take the ready-made one. The tool runs a single 4-stem separation — vocals, drums, bass, and other — and the 2-stem vocal/instrumental split is derived by summing the last three back together for you. That derived instrumental carries whatever artifacts the separation produced, same as the individual stems would, but grabbing it in one piece saves you the step of exporting three stems and balancing them back into a backing track by hand. If you'd rather keep the parts apart — to remix or sample them separately — the multi-stem splitter gives you all four with a live mixer.

Match the split to the job. A 2-stem result is the right shape for karaoke beds and acapellas; a 4-stem result is for remixing and sampling. Choosing more stems than you need doesn't make anything cleaner — it just gives you more seams to manage. When your task is to isolate vocals from a song, the 2-stem vocal is exactly what you want.

Clean up the residue afterward, gently. A lot of the watery texture lives in a narrow band. A gentle high-shelf tame or a light de-esser on the leftover fizz can push it below notice, especially on an acapella. Reach for a broad, subtle EQ move rather than aggressive surgery — heavy correction usually trades one artifact for another. When you're building a workflow to make an acapella from any track, this final polish is often the difference between usable and not.

Set expectations by genre. Before you spend an hour fighting a stem, check whether the material is even winnable. A sparse folk recording will clean up; a brick-walled EDM master won't fully, and that's the ceiling of the format, not a failure on your end.

A note on privacy while you experiment

Because all of this happens on your own device, you can throw as many source files and formats at it as you like without cost or a queue. The vocal remover runs one HTDemucs model (via onnxruntime-web) right inside your browser tab, so your audio never leaves your device. The one thing that does travel is the model itself: on first use it downloads to your browser from our CDN — tens of megabytes — and then caches for later runs. Separation isn't instant either; it works through your track on your own CPU, so give it real time on longer or denser songs. If you want the whole workflow end to end, the guide to split a song into stems covers it.

Frequently asked questions

Why does my separated instrumental sound watery or underwater? Two things stack up. Where a vocal and an instrument share the same spot on the spectrogram — a held note over a sustained pad, say — the model has to guess how to divide that shared energy, and the imperfect guess leaves a shimmering residue. On top of that, the parts of the process that work on the spectrogram reshape how loud each frequency is while leaning on timing (phase) closer to the original mix, which smears the audio slightly and adds that submerged, muffled quality.

Can I completely remove the artifacts? No, and any tool promising perfection is overselling. Wherever two sounds genuinely overlap in the same frequencies at the same instant, some ambiguity is mathematically unavoidable — even a flawless model has to split shared energy somehow. The realistic goal is to reduce the bleed: start from a lossless source, grab the ready-made instrumental instead of rebuilding it from stems, and gently EQ or de-ess the residue afterward.

Does a higher-quality source file actually help? Yes, noticeably. A low-bitrate MP3 has already discarded high-frequency detail and baked in its own compression artifacts, and the separator can't recover deleted information — it can even amplify the encoder's damage. A lossless WAV or FLAC, or at least a high-bitrate file, gives the model far more to work with, so the split comes out cleaner. It's the single most effective thing you can change.

Why do some songs separate cleaner than others? It comes down to how much the instruments overlap. A sparse arrangement leaves each sound its own room in the spectrogram, so the model rarely has to guess. A dense, loud, heavily limited modern master stacks everything into the same frequency space, creating far more contested tiles and far more residue. Reverb-heavy, doubled, or effect-drenched vocals spread across the mix and bleed more too.

Is the acapella from the vocal remover different from the stem splitter's? No — it's the identical stem. Both tools run the same HTDemucs model in a single pass; the stem splitter just shows you all four parts, while the vocal remover sums three of them into one instrumental. Switching tools won't give you a cleaner voice, because it's the same output. To reduce artifacts, change the input quality or the post-processing, not which tool you open.