Cleanup
Removing reverb from vocals: when it helps, and when to leave it
Reverb is the echo a room adds to a voice — the sound of the space, not the singer. An AI voice model cannot tell the two apart, so it copies the echo along with the person. That matters a lot when you are training a voice, and much less when you are simply converting one. There is also a third use for it that almost nobody knows about, further down.
This is the cleanup tool people reach for last, because echo does not sound like a fault. It sounds like the record.
What reverb actually is
When someone sings in a room you hear the voice twice: once directly, and again a split second later as it bounces off the walls. That second, blurry copy is reverb. Studios add it on purpose, because it makes a voice sound big and finished.
The problem for AI is simple. To a voice model the echo is just more sound arriving. It has no idea that part of it is a person and part of it is a wall.
Where it really matters: training a voice
A voice model learns whatever is the same across all your clips. If you recorded everything in one echoey room, that echo is in every single example — which makes it one of the most consistent things in the whole set, more consistent than any individual word.
So the model decides the echo is part of the voice. From then on it puts that room on everything you make, and there is no setting to turn it off later.
This is the one you cannot undo. A poor source vocal costs you one conversion. An echoey training set costs you every conversion that model will ever make. The same goes for background noise — anything running through the whole set gets learned as part of the singer.
Converting a vocal: often you can leave it
Converting the vocal exactly as it came out of the song, echo and all, works well a lot of the time. The space it was sung in is part of how the performance sounds, and taking it away can leave the result flat and dry against the music it has to sit next to.
It is a taste decision. If the echo is getting in the way, take it off. If the track sounds right, leave it.
The advanced trick: keep a clean copy to borrow from
This one takes a minute to explain, and it is useful once it clicks.
AI vocals are famous for losing their consonants — the sounds between the vowels, like s, t, k and p. (Here is why that happens.) The way to fix it is to take those sounds from the original recording, where they still sound right, and drop them into the converted version.
Which means you end up using the original recording for two different jobs:
- The copy you convert. Leave the echo on. It keeps the character of the performance.
- The copy you borrow sounds from. Take the echo off this one.
Why the difference? Because when you lift an s out of an echoey recording you do not just get the s. You get the s plus the little tail of room that follows it. Drop that into your converted vocal and you have added a splash of a room that exists nowhere else in the track. It sticks out.
Take the echo off that second copy first, and the borrowed sounds arrive clean. They sit in the new vocal without announcing themselves.
Worth knowing: the automatic repair in NiceVois borrows those sounds from the same vocal you uploaded, so it does not keep a separate clean copy. If your vocal has a lot of echo, some of it comes along. The two-copy method is a manual one — run the reverb remover yourself and do the swap in your own editor.
What to expect
Reverb removal reduces an echo; it does not delete it. A little room comes off cleanly. A recording made in a big hall, or a vocal with heavy studio echo printed onto it, is genuinely hard — and pushed too far, the tool leaves a thin, clipped sound that is worse than the echo was.
Trust your ears over the setting. If the cleaned version sounds unnatural to you, it will sound unnatural to everyone else.
Common questions
What is reverb?
The echo a room adds to a voice. It is the sound of the space rather than the person singing.
Should I always remove it before converting?
No. Many people convert with the echo still on, because it keeps the feel of the original. It is a choice about the sound you want.
Is it worse in training audio?
Much worse. One echoey vocal affects one conversion. An echoey training set affects everything that model ever makes.
What if the only recording I have is echoey?
Clean it as far as it goes without sounding strange, and accept the rest. A slightly echoey vocal works better than an over-processed one.
Ten free minutes, with its own allowance separate from the other tools.
Open the reverb remover