The No.1 Website for Pro Audio
 All  This Thread  Reviews  Gear Database  Gear for sale     Latest  Trending
convolution reverb removal
Old 21st February 2006
  #1
convolution reverb removal

I'm curious to know if there's any convolution software that you can use to take an impulse of a room and remove the reverb character from recordings after (or maybe even during) tracking. For the next month I'm living away from my studio and just have a small bit of equipment and would like to make the best of it. My room sounds terrible and I started thinking that I'm certainly not alone in that scenario. Seems like this "room subtraction" could be done to some degree but I've not aware of anything that could do it.
Old 21st February 2006
  #2
Lives for gear
 
Steamy Williams's Avatar
 

If it's possible you could really be on to something - that could be the best new idea for a product I've heard in a long time!
Old 21st February 2006
  #3
It seems like the technology is already there and is implemented in similar ways. All you'd really need is a mix contol to bring take more or less of the room out as there could be strange artifacts popping up at %100 removal. I'm unsure whether the impulse grab could be made with the same mic as subsequent recordings without effecting the sound of that microphone but if that were possible you could set up your mic in the placement that you're going to record in, snap your impluse, and have a custom room removal for that exact setup. Move the mic, snap another impulse. I don't know the science involved, but it sure would be interesting if it worked.
Old 23rd February 2006
  #4
so.. if nothing exists, can anybody think of a workaround? can any existing programs be used to this effect?
Old 23rd February 2006
  #5
Quote:
Originally Posted by aeroc
so.. if nothing exists, can anybody think of a workaround? can any existing programs be used to this effect?
Yep

Old 23rd February 2006
  #6
Lives for gear
 

well, you could mic the room and put those mics out of phase....you'd probably have to move them around a bit to get them to be 180 out.

just an idea...i've never tried it.

chris
Old 23rd February 2006
  #7
Lives for gear
 

If you could do this (you can't by the way) a lot of ****ty recordings could be made more listenable!

What you have not considered is that the reverberation is not only occuring after a note, but from the previous note DURING the note as well.
No algorythm could possibly distinquish information resulting from "bad" reverb from the "good" note.

It is all a stream of frquencies observed at a given moment, but our brain seperates it out into segments of time because of past experience with time.

It's kind of like un-mixing paint.

In fact, it is almost like attempting time travel if you really think about it!

You could possibly seperate stuff, but you'd destroy the stuff you intended to keep, too.

This idea/desire has been around ever since someone decided that the acoustics of a room sucked when they heard a recording played back.
Unfortunately, in the digital age people think that there might be some digital magic possible.

It is about as possible as removing the vocals from a recording.
It doesn't quite work does it?

Danny Brown
Old 17th March 2008
  #8
Here for the gear
 

Deconvolution is possible

Quote:
Originally Posted by dbbubba View Post
If you could do this (you can't by the way) a lot of ****ty recordings could be made more listenable!

What you have not considered is that the reverberation is not only occuring after a note, but from the previous note DURING the note as well.
No algorythm could possibly distinquish information resulting from "bad" reverb from the "good" note.

It is all a stream of frquencies observed at a given moment, but our brain seperates it out into segments of time because of past experience with time.

It's kind of like un-mixing paint.

In fact, it is almost like attempting time travel if you really think about it!

You could possibly seperate stuff, but you'd destroy the stuff you intended to keep, too.

This idea/desire has been around ever since someone decided that the acoustics of a room sucked when they heard a recording played back.
Unfortunately, in the digital age people think that there might be some digital magic possible.

It is about as possible as removing the vocals from a recording.
It doesn't quite work does it?

Danny Brown
There's nothing impossible about deconvolution in principle.

Reverb (to a first approximation) is a linear process. That means that the original sound and the reverberations are simply added together. The fact that the reverb stretches out over multiple notes doesn't make the math harder. Given the convolution the original went through, there's a deconvolution that will exactly reverse the process and give you the exactly dry signal. Heh.

There are only two problems. The first is getting the convolution function of the room. You can put a speaker where you think the source was, and put the mikes where you think they were, and make sure all the people, instruments, props, windows, air temperature, etc. are exactly the way they were, and record known clicks or white noise with reverb, and calculate a convolution function. But problem one is that there was probably more than one sound source, it wasn't a point source, and it wasn't holding still and neither was the room.

Then you invert the convolution function to get the deconvolution. It's equivalent to a graphic equalizer with 44100 or 48000 or 96000 bands for every second of reverb. Problem two is that some of those bands will have very high gain. If everything were perfect, they would all cancel except for the dry original. But because of noise and limited accuracy of digitization, you get the noise amplified through these hi-gain bands in a way that doesn't cancel quite so well anymore. What has happened is that some of the information that would have enabled perfect reconstruction has been pushed down beneath the noise and sensitivity of the recorder.

So, because reverb is basically linear, deconvolution is only "n log n"-- not nearly as hard as descrambling eggs. But because everything's imperfect, the problem is worse than just measuring and undoing the room's reverb function.

I have never heard the simple version of this technique tried, so I don't know how well or badly it does. There are complex heuristic algorithms to try to deal with the noise in deconvolutions, but I don't know whether they've been applied to this problem.

There could also be heuristics to deal with multiple, changing sources and changing rooms. Haven't found any so far.
Old 17th March 2008
  #9
Lives for gear
 

If anyone could do it, maybe the Celemony (Melodyne) guys could. If they can isolate notes in a polyphonic chord, and tune or shorten that individual note - what couldn't they do next?

I would have argued that was impossible to remove reverb - but maybe it isn't.

Our brain/mind has the amazing ability to be able to hear "into" mixed signals and filter out what we deem important. At some level, this is a form of data processing. As much as some would like to believe they understand how this all works, i'm convinced nobody has really got a clue.

So I would say it's possible in theory, because are brains are doing this all the time. If you are listening to somebody talking to you, walking around a hall, the mixture of reflections is constantly changing, but we subconsciously filter it all out and can follow the conversation effortlessly. We take it for granted because it's effortless - but at some level, there has to be some incredible number crunching going on. Or maybe it's pure analog, which would be more amazing.
Old 10th April 2011
  #10
Gear Head
 

You would do the opposite of creating an impulse. Take a sine wav with reverb and use it in Deconvolver as "unprocessed" file, and the regular unprocessed sine wav as the "processed" file. Worth a shot...
Old 10th April 2011
  #11
Gear Nut
 
DITN's Avatar
Or you could try the SPL De-Verb.
Old 10th April 2011
  #12
Gear Nut
 
The Fold Studios's Avatar
 

Quote:
Originally Posted by dbbubba View Post
No algorythm could possibly distinquish information resulting from "bad" reverb from the "good" note.
Never say never. A lot of people were rubbishing Melodyne's DNA technology before it came out (I'm sure you could find plenty of posts on here by well informed people a couple of years ago saying that being able to seperate and manipulate componant notes from a single polyphonic audio file is impossible).

Apparently Melodyne are now working on technology to remove unwanted microphone spill so...
Old 10th April 2011
  #13
Gear Head
 

is this the sort of thing you looking for?

ARC System
Old 10th April 2011
  #14
Lives for gear
 

Quote:
Originally Posted by The Fold Studios View Post
Never say never. A lot of people were rubbishing Melodyne's DNA technology before it came out (I'm sure you could find plenty of posts on here by well informed people a couple of years ago saying that being able to seperate and manipulate componant notes from a single polyphonic audio file is impossible).

Apparently Melodyne are now working on technology to remove unwanted microphone spill so...

Thats like asking for a pill to cure unhappiness without side effects

Sure, melodyne can be a nice life but it doesn't sound very good.
Old 10th April 2011
  #15
Lives for gear
 

In principle the deconvolution idea is possible, but putting it into practice and achieving decent results are very tricky.

In order to remove the reverb from a given recording, an IR would need to be made in the exact same room using the exact same mic, with the room containing the exact same things in the exact same places ... and people ... etc ...

In theory you could then use that IR to remove/attenuate the existing room tone from that one mic signal ... however I highly doubt that the results would be particularly good.

... basically just repeated FutureNerd
///

The only thing I've ever heard that can remove/attenuate reverb is the CEDAR DNS ... that works pretty well from the examples I've heard ... but only on voices as far as I'm aware, as that is what it's designed for.

Last edited by -tc-; 10th April 2011 at 03:32 PM.. Reason: what is up with my spelling today ?!!
Old 10th April 2011
  #16
Lives for gear
 
travisbrown's Avatar
Quote:
Originally Posted by not_so_new View Post
Okay. Tell me how you use a Transient Designer to remove reverb. I posed the question of removing reverb a while ago here when I got some live-recorded tracks that had way to much hall sound in them.
Old 10th April 2011
  #17
Lives for gear
 
MickeyMassacre's Avatar
Why not spend a few hundred bucks making the room not sound so bad?
Old 9th August 2012
  #18
Here for the gear
 

You most certainly can.

Bare in mind though that a true impulse is a mathematical construct and does not exist in reality. Impulse room sampling is only a rough approximation. There are other methods of deconvolution that depict a more accurate transfer function such as using frequency sweeps but those require substantially more complex algorithms.

Convolution is simply:

y[n] = h[n]*x[n]. (the star means convolution, not multiplication).

A property of convolution is to convert h[n] and x[n] to functions in the frequency domain using the Fourier transform. You then have the property:
Y[jw] = H[jw]X[jw] (simple multiplication). From here you can clearly see that you can solve for either X (your impulse in frequency domain) or H (your room reverb).

Next you preform a Fourier series inverse and you have your room characteristics: x[n]. This function x[n] can be convolved with arbitrary signals or used to deconvolve an identical x[n] out of a signal.
Old 9th August 2012
  #19
Lives for gear
 

I'd just use Zynaptiq Unveil. It didin't exist 7 years ago when this thread started, but does now,
Old 10th August 2012
  #20
Here for the gear
 

Quote:
Originally Posted by thenoodle View Post
I'd just use Zynaptiq Unveil. It didin't exist 7 years ago when this thread started, but does now,
Point being, it's possible. The whole reason I responded in the first place was to dispel any doubt made by some of the commenters. I was searching the web for information on Fourier transform sweeps and came across this thread. Everything that program does could have been done in MATLAB when this thread was started with about 20 lines of code. Anyway, happy mixing folks.
Old 10th August 2012
  #21
Lives for gear
 
Denis Goekdag's Avatar
 

No, what UNVEIL does is not possible in MATLAB with 20 lines of code. IF you have the IR from the room the recording was made in, and IF that IR is very accurate (sweep speaker placed at exactly the same position as the recorded source, mic placement exactly the same), then yes, you could remove reverb using convolution methods in MATLAB.

But only for that particular room, with that particular setup, and only if you have the IR.

UNVEIL, on the other hand, removes reverb from any recording, without having an IR. It's a "blind" process, which means it doesn't require additional information about the signal (like an IR) to do what it does.
Old 10th August 2012
  #22
Here for the gear
 

Quote:
Originally Posted by Denis Goekdag View Post
No, what UNVEIL does is not possible in MATLAB with 20 lines of code. IF you have the IR from the room the recording was made in, and IF that IR is very accurate (sweep speaker placed at exactly the same position as the recorded source, mic placement exactly the same), then yes, you could remove reverb using convolution methods in MATLAB.

But only for that particular room, with that particular setup, and only if you have the IR.

UNVEIL, on the other hand, removes reverb from any recording, without having an IR. It's a "blind" process, which means it doesn't require additional information about the signal (like an IR) to do what it does.
Ultimately it is deconvolution. That is the only way you can remove a reverb from a mix. The program very likely analyses the frequency information using the Fourier transform and is able to identify dominant harmonics that might be the culprit then gives you a rough estimate of what the transfer function should be. From there you can make adjustments to your liking. Now we're at 30 lines of code. By 30 lines of code I mean using built in functions in MATLAB that handle this stuff that are much larger packages than the code required for their implementation. Those packages also exist because these theories have existed for quite a while and were used in communications decades before anyone thought of using them in a studio. That said, my comment on the original post still holds true, it's possible. Granted, the processing would be slower and there would be no GUI. Just saying. Happy mixing.
Old 11th August 2012
  #23
Lives for gear
 
Denis Goekdag's Avatar
 

Dude...I *make* UNVEIL, so I believe I have some idea of what it does....heh

It doesn't do any (de-)convolution, and it doesn't use any FFT.

Just sayin' ;-)
Old 11th August 2012
  #24
Lives for gear
 

Is Unveil staying Mac-only?
Old 11th August 2012
  #25
Lives for gear
 
Denis Goekdag's Avatar
 

Quote:
Originally Posted by thenoodle View Post
Is Unveil staying Mac-only?
No :-) We're working on VST and PT support for both Mac and Win, coming sometime this fall/winter. This will be a free update for existing users.
Old 12th August 2012
  #26
Here for the gear
 

Quote:
Originally Posted by Denis Goekdag View Post
Dude...I *make* UNVEIL, so I believe I have some idea of what it does....heh

It doesn't do any (de-)convolution, and it doesn't use any FFT.

Just sayin' ;-)
Then please explain to me theoretically how it "removes" a convolved signal without using deconvolution. I'm simply speaking from a perspective of a person about to earn an MS in signal processing. Once again, all systems that apply a reverb, be it digital, electrical or acoustic are represented by this equation in the frequency domain: Y[jw] = H[jw]X[jw]. If you can somehow break the laws of physics and rules of mathematics to solve for X you are a wizard. If your program achieves this without using convolution methods I'd say you folks are in the wrong industry. I'm going to assume if you designed this system you can at least hint to me the methods that you use. Otherwise if you invented any new theoretical techniques for removing components from a convolved signal, you would be gaining world engineering recognition and probably wouldn't be selling products to recording artists, and also, in that case, please give me the link to the patents that you surely hold.

In case anyone was wondering, this is how long this stuff has been around: http://en.wikipedia.org/wiki/Joseph_Fourier

And here's some more well defined terminology: http://www.wisdom.weizmann.ac.il/~le...EtalCVPR09.pdf

http://perso.telecom-paristech.fr/~cappe/dcv/ (notice the date on this one)

Just looked at some of your posts, and the link http://www.zynaptiq.com. It appears that you folks bought rights to use existing code packages. I'm going to go out on a limb here and say that no, you don't know how your software works.
Old 13th August 2012
  #27
Lives for gear
Quote:
Originally Posted by Denis Goekdag View Post
Dude...I *make* UNVEIL, so I believe I have some idea of what it does....heh

It doesn't do any (de-)convolution, and it doesn't use any FFT.

Just sayin' ;-)
Hey Denis, nice to see you weigh in here. I got a chuckle on the *make*.
Old 13th August 2012
  #28
Here for the gear
 

Quote:
Originally Posted by Syncamorea View Post
Hey Denis, nice to see you weigh in here. I got a chuckle on the *make*.
Lol, what makes me chuckle is that I have to explain fundamental signal processing theory to a guy who runs a company that has the slogan "Science, not fiction".

Don't worry Dennis, no one here is going to use MATLAB to solve this problem. Your program will set someone back $399 whereas MATLAB will set someone back 10 grand.

The advantage of MATLAB is it's a well established product that is the de facto standard in signal processing research and has been around since 1984. We'll see how well your company fares, but I wish you the best.

All that said I assure you, I can remove reverb in MATLAB, *I'm an engineer, I might know how MATLAB works*. I perform deconvolution all the time.

In fact, wireless communications would not be possible without the very thing your program does.
Old 13th August 2012
  #29
Lives for gear
I just bought two Matlab licenses for about half that, IIRC.

Minor point, I admit. However, plugging Matlab into Cubase makes my monitors warble and into Logic, it makes Spock sound illogical. Maybe I need to buy new presets.
Old 13th August 2012
  #30
Here for the gear
 

Quote:
Originally Posted by Syncamorea View Post
I just bought two Matlab licenses for about half that, IIRC.

Minor point, I admit. However, plugging Matlab into Cubase makes my monitors warble and into Logic, it makes Spock sound illogical. Maybe I need to buy new presets.
They do have different limitations for how their software can be used. For example, instructional or educational purposes are significantly discounted but you likely wont own anything you create from those licenses.

I dont think MATLAB can be used in that sense but I've never looked into it. I understand it as simply a programming platform for science and engineering. It's a great tool though for experimentation. I've processed some very strange effects in it before. Pretty fun if you ask me.

Edit:
Yes upon further inspection I stand corrected on the pricing. Some folks are saying a commercial copy is $1900. I often use it at the U and I have a student edition at home. Well worth the time and money though I'd say. I wouldn't mind learning cSound as well.

I'd say probably the best thing to use for reverb removal would be Denise's program though. I originally was only trying to make a point for all the naysayers. To deconvolve a reverb out of a mix in Matlab on two channels takes about half hour to compute on a 2.4Ghz dual core processor. Not exactly ideal. Most Universities have supercomputers that spit out results from Matlab commands months later.
Post Reply

Welcome to the Gearslutz Pro Audio Community!

Registration benefits include:
  • The ability to reply to and create new discussions
  • Access to members-only giveaways & competitions
  • Interact with VIP industry experts in our guest Q&As
  • Access to members-only sub forum discussions
  • Access to members-only Chat Room
  • Get INSTANT ACCESS to the world's best private pro audio Classifieds for only USD $20/year
  • Promote your eBay auctions and Reverb.com listings for free
  • Remove this message!
You need an account to post a reply. Create a username and password below and an account will be created and your post entered.


 
 
Slide to join now Processing…
Thread Tools
Search this Thread
Search this Thread:

Advanced Search
Forum Jump
Forum Jump