Login / Register
 
Can an impulse response be used in reverse?
New Reply
Subscribe
juicylime
Thread Starter
#1
21st June 2007
Old 21st June 2007
  #1
Lives for gear
 
juicylime's Avatar
 
Joined: Sep 2006
Location: Dublin, Ireland
Posts: 809

Thread Starter
juicylime is offline
Can an impulse response be used in reverse?

I've kind of being wondering about this and perhaps it's a stupid question, but I'd like to know. If you make an impulse response, e.g. an electric spark in a cathedral, can you then somehow (and I'm talking theoretically here) take the original impulse response recording and actually use the data of the impulse response to remove the sound of the cathedral so you are left with just the sound of an electric spark and no cathedral sound? A bit like reversing the phase of the IR (but in another way not all, right?), this could have some pretty damn useful applications if you could. I know you maths guys are out there!!
__________________
Saved. By the buoyancy of citrus.



#2
21st June 2007
Old 21st June 2007
  #2
Lives for gear
 
Joined: Feb 2006
Posts: 4,075

Kiwiburger is offline
What you are really asking for is a way to remove reverb from a sample. There is no really successful way to do this. In your example, you would be far better off simply recording your electric spark in an anechoic chamber - and then using your choice of reverb later.

To a limited extent, you can remove reverb by using Mid/Side processing. Turn the sample to mono, and the centre/dry information gets a 6dB boost from phase coherent summing, and the side information partially cancells, or at least does not get the same amount of boost.

After that, an envelope modifier (e.g. a fade-out) and possibly Expansion can remove the reverb tail.

If you had wet and dry samples of the exact same audio event, then you wouldn't need any special processing because you already have the dry sample. So I can only assume you have a wet sample and want to extract out the dry sound - and that isn't easy.

This is sort of another re-statement of the old Karoke vocal-removal question that gets asked often. I believe there is some special software that can do this better than the above techniques. It works on the FFT process, and by looking at stereo pan position. But I think it's still pretty dodgy, because - say for example - you have a 100 Hz sinewave dry signal. The reflected sound coming back will basically only contain 100 Hz, just multiple reflections. You can't really notch it out with eq, because that would notch out the dry sound. You could look at the differences between left and right channels, and assume anything different is the reflection and therefore drop it out. But then you are pretty much back to Mid/Side technique.

You can't unscramble an egg.
__________________
My carbon footprint is bigger than yours.
#3
21st June 2007
Old 21st June 2007
  #3
Gear Guru
 
tINY's Avatar
 
Joined: Jul 2004
Location: Orygun
Posts: 12,082

tINY is offline


Yes, you can do it.

But, you have to use the same mic in the same position and the cathederal should be almost identical to the time that you recorded the impulse.

Oh, yeah .... The sound source needs to be in the same place as the sparker and the temp and humidity need to be the same as when the sparker was recorded.

How well does it work in the real world? I think the Kiwi gave the correct practical answer above.





-tINY

juicylime
Thread Starter
#4
21st June 2007
Old 21st June 2007
  #4
Lives for gear
 
juicylime's Avatar
 
Joined: Sep 2006
Location: Dublin, Ireland
Posts: 809

Thread Starter
juicylime is offline
Kiwiburger: The question isn't actually about getting rid of reverb at all, I just used that as an example. It could as much be an IR of a Neve pre-amp or a distortion pedal. The question is can the data gathered be used, not just to re-create a certain sound, but also to remove it?

tINY: An exact recording would exist, the one used to create the actual IR in the first place. Can you explain to me how it would then work?
#5
21st June 2007
Old 21st June 2007
  #5
Lives for gear
 
John Suitcase's Avatar
 
Joined: Jul 2005
Location: Phoenix, AZ
Posts: 2,397

John Suitcase is offline
Search around on Google for deconvolution and 'blind deconvolution.' This is one of the holy grails of communications technology, since it would allow much cleaner signal to pass through less robust noisy networks.

Theoretically it should be possible, but as far as I know, no one has done it with the amount of success you're looking for.

Some things, like a distortion pedal, are pretty lossy from an information perspective. The waveforms are being clipped, flat-topped, so any 'data' that existed in that part of the waveform would be lost. Some high-frequency stuff, for instance.

If you know what the original signal 'probably' looked like, you could deconvolve the distorted signal and get something close to the original undistorted signal. That is, if you know that it was a sine wave, for instance, you can easily get the original waveform from one that's been heavily clipped. As the waveform gets more complex, it gets harder.
#6
21st June 2007
Old 21st June 2007
  #6
Gear Guru
 
tINY's Avatar
 
Joined: Jul 2004
Location: Orygun
Posts: 12,082

tINY is offline


When you take an IR, the computer looks at the resulting waveform and constructs an FIR filter.

This filter has a bunch of delays with a co-efficient value for each "tap". For 44.1k sampling, these taps are offset from each other by about 22.676 microseconds. To convolve a signal with the impulse response, the signal is sent through the FIR, which sends it out and adds it with the delayed signals multiplied by the co-efficients.

You can reverse this process mathmatically - there will be a long delay from input to output (equal to the longest tap delay in the initial convolution).



-tINY

FDI DSP - Digital Signal Processing - Design Guide
Finite impulse response - Wikipedia, the free encyclopedia
Convolution - Wikipedia, the free encyclopedia
Deconvolution - Wikipedia, the free encyclopedia



#7
21st June 2007
Old 21st June 2007
  #7
Lives for gear
 
max cooper's Avatar
 
Joined: Aug 2004
Location: tx
Posts: 8,803

max cooper is offline
Quote:
Originally Posted by tINY View Post
But, you have to use the same mic in the same position and the cathederal should be almost identical to the time that you recorded the impulse.
Kinda o/t, but you can do a similar thing when you record a singer in the midst of a live band; you have to have two of the same microphone in similar positions but one has to be out of capsule-shot of the singer, more or less.

In other words the mics should hear about the same thing except for the singer, then you subtract one from the other by using polarity.
#8
25th January 2009
Old 25th January 2009
  #8
Gear interested
 
Joined: Jan 2009
Posts: 10

BobChutney is offline
I know this is an age-old thread. I thought I'd revive it rather than starting a new one.
My brother, who is not a sound engineer but a geophysicist, tells me this should be possible, and that he uses deconvolution all the time in analogous ways. I work quite a bit with a singer who records his vocals through a cheap condensor mic in a Turkish cottage. So what my brother says to me is that yes, Bob, you can create a sort of un-cheap condenser and turkish bedroomizer. He's running something on a super-computer for me now. I'm not holding my breath.
#9
25th January 2009
Old 25th January 2009
  #9
Moderator
 
narcoman's Avatar
 
Joined: May 2007
Location: UK
Posts: 11,734

narcoman is offline
Quote:
Originally Posted by BobChutney View Post
I know this is an age-old thread. I thought I'd revive it rather than starting a new one.
My brother, who is not a sound engineer but a geophysicist, tells me this should be possible, and that he uses deconvolution all the time in analogous ways. I work quite a bit with a singer who records his vocals through a cheap condensor mic in a Turkish cottage. So what my brother says to me is that yes, Bob, you can create a sort of un-cheap condenser and turkish bedroomizer. He's running something on a super-computer for me now. I'm not holding my breath.
but your brother uses deconvolution in a way to provide a statistical average information return based on predictability. What sound engineers want isn't a "predictable" result - but the original EXACT result. In this scenario you need the original information.

The reason it doesn't work in exact reductive circumstances is provable via induction. When using a convolution reverb or IR you are using a static model (or pseudo random seeded model) to provide apparent uncorrelated results. That very random nature (even if seeded) stops you from "deconvolving" since there is an uncorrelated output. For deconvolution to work you need correlated or at least statistically predictable correlation.

Not doable for sound or any other uncorrelated signal. You MIGHT get useful and pleasing results, but because of the random and uncorrelated nature you would never be sure of the original signal without having the original signal to compare...
#10
25th January 2009
Old 25th January 2009
  #10
Lives for gear
 
Joined: Feb 2006
Posts: 4,075

Kiwiburger is offline
The intriguing thing is that our brains decode dry sounds from reverb sounds all the time. If we listen to somebody walking around a room, following their conversation, we aren't overwhelmed by the constantly changing ambiant reflections. We seem to be able to ignore them and focus on the words - and we can recognise people and instruments coming at us from totally different spaces all the time. The reflections are processed - we can identify locations and other clues from them - so somehow our brains are making a lot of sense from the data in a way that current digital processing can't approach. The 'software' supporting our brains is mind blowing to contemplate ... and frankly I don't believe it is biological or totally within the four known dimensions.

Deconvolution could possibly work if the reflections were generated by a precisely repeatable process. But real spaces, and many digital spaces, have random elements that would make this impossible.

Anyhoo ... although I don't think deconvolution is the answer, I wonder if resynthesis might provide a better method? If a sound is reconstructed out of sinewave partials, then maybe we could use logic to weed out the partials that seem to be have been created by reflections rather than the dry source? Probably easier with a stereo file - we could assume that data that was not exactly common to both sides was from the reflections ...

I don't know - it seems like a highly complex problem that won't be solved in our lifetime, but if our brains can process this data, there probably is a way ...
#11
25th January 2009
Old 25th January 2009
  #11
Lives for gear
 
Joined: Apr 2004
Posts: 1,230

NeoVXR is offline
I think lack of knowledge of the "original", in engineering often can be replaced by knowledge of the desired result!
eventually most time what you want is a pleasant result, not metaphysical truth..


the human ear (plus brainware) can understand a reverberated voice only when it has heard so many clear voices already. there is a factor of prediction and expectance.
__________________
sorry 4 poor english
#12
25th January 2009
Old 25th January 2009
  #12
Lives for gear
 
Joined: Feb 2006
Posts: 4,075

Kiwiburger is offline
Yeah - it's 'fuzzy logic' and parallel processing of some description ... that's why I think resynthesis might be closer to what is required ...
#13
25th January 2009
Old 25th January 2009
  #13
Gear maniac
 
Joined: Jun 2006
Posts: 206

Soulbrother is offline
It would theoretically be possible

But remember that whatever method you use to simulate the impulse response (balloon pop, starter pistol, handclap) will all affect the frequency response of the impulse, so it will never be a 'perfect' impulse of the room.
#14
25th January 2009
Old 25th January 2009
  #14
Lives for gear
 
John Suitcase's Avatar
 
Joined: Jul 2005
Location: Phoenix, AZ
Posts: 2,397

John Suitcase is offline
Quote:
Originally Posted by Soulbrother View Post
But remember that whatever method you use to simulate the impulse response (balloon pop, starter pistol, handclap) will all affect the frequency response of the impulse, so it will never be a 'perfect' impulse of the room.
What about if you use a known impulse, say send him a clicker or other device which you have already samples at your location? I agree it wouldn't be 'perfect' but it should be close enough to improve your results significantly.

I'm no math wiz, but it seems like you could probably use the IR info to get back to something close to the original sound minus any artiacts from the room or mic (Clearly it's difficult if not impossible to recreate frequencies that are completely gone from the vocal recording. You couldn't use an impulse of the room recorded using a subkick to then extrapolate the high frequency information of the sung vocal that was also recorded using the subkick. But as long as you're simply trying to remove unwanted information (the room, maybe some over-emphasized frequencies from the mic) it should work, I would think.
New Reply Submit Thread to Facebook Facebook  Submit Thread to Twitter Twitter  Submit Thread to LinkedIn LinkedIn  Submit Thread to Google+ Google+ 
 
Thread Tools
Search this Thread
Search this Thread:

Advanced Search
Similar Threads
Thread
Thread Starter / Forum
Replies
Joseph Reeves / Rap + Hip Hop engineering & production
1
lyricstree / Music Computers
3
The Vicar / So much gear, so little time!
7
ultima / So much gear, so little time!
37

Forum Jump

SEO by vBSEO ©2011, Crawlability, Inc.