Q:could MConvolutionMB be used to REMOVE reverb?

[steve2KVR]

I've experimented with several ideas/etc and nothing worked, and I'm not at all convinced the approach has any merit. Just asking about some possible solutions...Cheers!.../s~

[vectorwarrior]

Convolution adds content to a signal, how do you propose that this would work? I'm guessing some sort of phase cancellation, but it's not clear how on earth you'd do that. What's your idea?

[imrae]

I think the process you're after is https://en.wikipedia.org/wiki/Deconvolution

The maths is different from convolution, you can't just do it by modifying the kernel.

I'm not aware of any plugin that applies this for music. Might be interesting, but I'd guess it's pretty noisy in practice.

[steve2KVR]

It's not so much and "idea" as working out a solution to typical problem. Currently everything I'm working was done in someone's basement/garage/bedroom/etc and so the recording space is what it is, that's what they've got. I've got time, they've money. So my thought was if I can add "reverb/resonance" with an IR of a room, maybe I could remove the resonance/reverb as well.

I tried all of the de-verbs on the market and the results are just so-so when it comes to instruments, particularly anything legato. Staccato/spoken-word seems to work the best.

So as a thought experiment it seemed on the face of it that if I can create just about any space and add it in, maybe I could create spaces that could be taken out. Alone phase reversal does next to nothing as far as "removal" is concerned.

I'm open to any suggestions anyones got...Cheers~.../Steve~

[imrae]

The theory is that if you know exactly what impulse response was applied to a signal (i.e. you have an impulse from the same source for the same mic setup) then you should be able to mathematically find a model dry signal that would have give the same result as your recording after convolution with that impulse. The problem is that if your impulse is even slightly wrong, the results of this fitting might be quite strange. Would need to try it to find out, but I'm confident that a broadly-similar-sounding reverb impulse is not going to cut it.

[steve2KVR]

@imrae: My intuition is that in the details you are entirely correct, the complexity of the maths is probably quite extreme. And the situation is made even more difficult in that I don't have an initial IR for the spaces. It's just the recordings. I've requested pictures/etc. and created estimated spacial dimensions, and from those theorized the likely eigentones of the space to reverse engineer some possible reverb/resonance responses...meh...thanx for the thoughts.../s~

[imrae]

Thinking about it, the software used for generating IRs is actually a deconvolver. Convolution is a commutative operation, it doesn't care which audio is the "input" and which is the "kernel", the result should be the same. So for the reverse operation, we should be able to provide either input and find the other one...

I just tried this in REAPER and it worked!
- clean 808 snare sample
- add convolution reverb with "jazz hall" IR, record to new sample
- open ReaVerb, add a File and access the "deconvolve" tool (usually used to make an IR from a sine sweep)
- select the snare+reverb sample as input
- select the "jazz hall" IR as the "test tone"

The output file is a clean 808 snare

[steve2KVR]

@imrae: NOW we're talkin'! I hadn't thought of that one! I shall proceed along those lines, and follow up...thanx.../s~

[vectorwarrior]

Ha ha, that's an interesting idea to reverse deconvolve like that! However, be aware that it will only work with a single hit. If you put a track or stems in, it will not work... or at least I would strongly imagine anyway...

[steve2KVR]

...fortunately I don't think I need perfect, just better/mostly-gone will do, and yes legato is the challenge.../s~

[spencerlee]

I think you are pissing in the wind. I don't think it is possible to do it that way as the acoustic space will be continually changing. Just the body movements will change the space. Then you will get that horrible phasing sound.
I have found in my 50 odd years of experience it is best to gate the shit out of the signal and add back a short acoustic space around the result.
You could try the gate technique but just gate to say -10db to reduce the reverb but leaving some of the signal intact.
Spencer

[steve2KVR]

@spencerlee: I'm not so pessimistic, after all I only need a pretty good result not mathematical precision. I'm looking to remove as much of the original acoustic environment as useful, and as you observe, add back a selected "synthetic" one later on.

After much experimentation using ARC4, REW, and SF4 I came to the conclusion that my being in the room in any location or not had no impact on the measured response. But did impact the listening experience. I have never experienced the phasing effect you mention.

I have been using a "dual channel out of phase compression gating" technique for quite sometime. Works best when including MADEq with dynamic filters approximately set to the room's eigentones. None of it is set and forget...so I was just asking how others have been solving the problem...Cheers!.../s~

[Faiky]

Once there is the slightest deviation in the IR, the technique will not work.

You can use MSpectralDynamics to remove reverb.
The noise profile is the ImpulseResponse, but you can also appreciate it, mostly the mids are louder and you are drawing a mountain in the spectograph.
Then you activate the gate function, as for denoising, and adjust the threshold. Then the fine adjustment of the other controllers like smooth and so on.

[steve2KVR]

Thanx for the suggestion, I'll give that a try in combo with some other thoughts...Cheers!.../s~

[MeldaProduction]

The problem with reverb removal is first that the reverb is "changing" and it technically masks the existing information in a destructive way. Just imagine a sine pulse train. Once you process it through reverb, it "glues" the pulses and nonmodulated echoes won't even change the frequency, so after that it's impossible to find out whether the glue is reverb or the original sound.

So gating could help (whether broadband or spectral), but at the end the only thing we can do is to sort of fool our brains. Most successful reverb removal works by using neural networks mimicking the typical scenarios in spectral domain. Hoopefully I'll get to that some day too, there's a lot once could get using these black boxes, but it requires dev time