Graphical sound denoising challenge

[A_SN]

I (over at Photosounder.com) have just started a new challenge. It consists in denoising a 115 year old sound by editing its image generated via Photosounder using Photoshop/GIMP/whatever you got or by writing a dedicated image processing algorithm.

The winners get a free license of Photosounder worth $140/€99.

Link to the contest with all the details, examples and files : Photosounder: Graphical sound denoising challenge

[Alexey Lukin]

Looks like you made a mistake: your daisy_bell_full.flac is already denoised, compared to the fragment you are presenting on the web-page.

[A_SN]

Originally Posted by Alexey Lukin

Looks like you made a mistake: your daisy_bell_full.flac is already denoised, compared to the fragment you are presenting on the web-page.

Fixed, thanks for letting me know by the way!

[wado1942]

Honestly, I don't really get it. Why would I do these confusing conversions with proprietary software that won't let me save my work when I can do the same job in under 5 minutes with 8+ year old audio software I already own?
http://gcmstudio.com/audioonly/daisy_bell_denoise.wav

[A_SN]

Originally Posted by wado1942

Honestly, I don't really get it. Why would I do these confusing conversions with proprietary software that won't let me save my work when I can do the same job in under 5 minutes with 8+ year old audio software I already own?
http://gcmstudio.com/audioonly/daisy_bell_denoise.wav

That's fine if you're satisfied with what it gives you. After hearing your example I wouldn't be satisfied with that though.

And what do you mean by "won't let you save your work"? Photosounder has a big Save button, so you can save, even in the demo you can save your work... And you can do the job under 5 minutes too :D. And what software do you use that isn't "proprietary"?

[It'sJoeAgain]

i don't know who made the example processed with ur software but imo the result sounds very good to me...

Edit; how do u get to filter out the pops and low crackles from the recordin without considerably thinning out the voice??

Oh don't tell me...spectrum analysis.. u can put put filters exactly where needed

[A_SN]

Originally Posted by It'sJoeAgain

i don't know who made the example processed with ur software but imo the result sounds very good to me...

I did it, and thank you. Actually that was my very first and only try with it. A friend of mine thought his own tests with a high pass filter of sorts and a seam carving thing sounded better though.. Well, the winner of the challenge will necessarily sound better anyways ;).

Edit; how do u get to filter out the pops and low crackles from the recordin without considerably thinning out the voice??

Well, basically with what I did in Photoshop I isolated the background (can't detail the operation too much before the challenge is over, but basically I used a bunch of copies of the same layer and did some operations on them and between them), and then did the difference, so in the background the pops and crackles stayed with it. That's exactly the kind of reasons why you want to Photoshop a sound, cause you can see what you want to get rid of then with a bit of thought you can come up with a way to Photoshop what you don't want out ;). The biggest problem I can see with my example is there's still a bunch of dots remaining, but I'm sure there's a way to Photoshop these out too!

[It'sJoeAgain]

Originally Posted by A_SN

I did it, and thank you. Actually that was my very first and only try with it. A friend of mine thought his own tests with a high pass filter of sorts and a seam carving thing sounded better though.. Well, the winner of the challenge will necessarily sound better anyways ;).

Check my edited comments above....another thing...i think the contest is a bit compromised by your example because your sample sounds really good, and the actual bandwidth or frequencies of interest are around 300Hz and 6K maybe 8Khz...so one doesn' have a lot of bands to work with and any *better* results one can get out with would be considered highly subjective imo...

[A_SN]

Originally Posted by It'sJoeAgain

Check my edited comments above....another thing...i think the contest is a bit compromised by your example because your sample sounds really good, and the actual bandwidth or frequencies of interest are around 300Hz and 6K maybe 8Khz...so one doesn' have a lot of bands to work with and any *better* results one can get out with would be considered highly subjective imo...

I don't know, I'm sure it can get better really. My example is far from perfect, and I know it can be perfected.

[It'sJoeAgain]

Originally Posted by A_SN

I don't know, I'm sure it can get better really. My example is far from perfect, and I know it can be perfected.

You have significantly reduced with filters the crackles and pops and all the weird friction artifacts in the recording, sure there are some phase issues as a result of it but a good trade off for the reduction of unwanted noise. But then what?? More filtering and you are cuttin' out the voice fundamentals.... so you can make it thinner and reduce the noise from the low middle range or mufflled by filtering out above 1.5kHz..again IMO the sample is balanced and as good as it's gonna get without compromising the actual fundamentals of the original recorded voice....

[A_SN]

Originally Posted by It'sJoeAgain

You have significantly reduced with filters the crackles and pops and all the weird friction artifacts in the recording, sure there is some phase issues as a result of it but a good trade of for the reduction of unwanted noise. But then what?? More filtering and you are cuttin' out the voice fundamentals.... so you can make it thinner and reduce the noise from the low middle range or mufflled by filtering out above 1.5kHz..again IMO the sample is balanced and as good as it's gonna get without compromising the actual fundamentals of the original recorded voice....

Phase issues? Not sure what that is, but it sounds like you're thinking in equalisation terms. It's very different, and to understand it, you really have to think in terms of image. Imagine that someones come up with the perfect ideal image representation of what the sound should be. Well you'd apply it with Photosounder onto the original sound and it would just work.

Of course no one's going to come up with such an image because there are some things that have been lost definitely to the noise, but you have to understand that it's all an image processing problem. If you can fix the image, you've fixed the sound. And you can fix the image without losing any of the voice components. You can potentially keep all that belongs to the voice, lose all that's from the noise, and eventually get to fix some of what was damaged by the noise. It can be done, I'm not sure how, but the goal of the whole challenge is that someone finds out. And that applies not just to denoising but pretty much anything. Fix the image and the sound will follow ;).

Also, I have to stress that it was just messing with an idea in Photoshop for like 5 minutes. Surely unless I'm incredibly lucky one can easily top that by putting more research into it!

[It'sJoeAgain]

Originally Posted by A_SN

Phase issues? Not sure what that is, but it sounds like you're thinking in equalisation terms. It's very different, and to understand it, you really have to think in terms of image. Imagine that someones come up with the perfect ideal image representation of what the sound should be.

ur talkin about phase distortion which btw it is added whether u realize that or not....i was referring to the recording that i am sure it was mono originally then converted to stereo by u...u don't notice the teeny phase shifting issues after the process??

Well you'd apply it with Photosounder onto the original sound and it would just work.
Of course no one's going to come up with such an image because there are some things that have been lost definitely to the noise, but you have to understand that it's all an image processing problem. If you can fix the image, you've fixed the sound. And you can fix the image without losing any of the voice components. You can potentially keep all that belongs to the voice, lose all that's from the noise, and eventually get to fix some of what was damaged by the noise. It can be done, I'm not sure how, but the goal of the whole challenge is that someone finds out. And that applies not just to denoising but pretty much anything. Fix the image and the sound will follow ;).
Also, I have to stress that it was just messing with an idea in Photoshop for like 5 minutes. Surely unless I'm incredibly lucky one can easily top that by putting more research into it!

Not tryin to be confrontational but IMO, me thinks you are trying to reconstruct the severely damaged sound recorded on a 19th century format and expect to make it sound like commercial broadcast quality...it's just not gonna happen 'cause the fundamentals for a clean voice in a quite room environment are simply not there and (been told by another engineer) that no spectrum analysis/imaging software will allow you to recreate the missing harmonics only cut or filter them out to preserve what sounds coherent with a certain degree of taste....That is to say whoever creates something more pleasing to the ears will win your contest....that's not improving the signal up and above what the technology or the tools can allow but a subjective judgment call....

[A_SN]

Originally Posted by It'sJoeAgain

ur talkin about phase distortion which btw it is added whether u realize that or not....i was referring to the recording that i am sure it was mono originally then converted to stereo by u...u don't notice the teeny phase shifting issues after the process??

It's all mono..

Originally Posted by It'sJoeAgain

Not tryin to be confrontational but IMO, me thinks you are trying to reconstruct the severely damaged sound recorded on a 19th century format and expect to make it sound like commercial broadcast quality...it's just not gonna happened 'cause the fundamentals for a clean voice in a quite room environment are simply not there and (been told by another engineer) that no spectrum analysis/imaging software will allow you to recreate the missing harmonics only cut or filter them out to preserve what sounds coherent with a certain degree of taste....That would be whoever creates something more pleasing to the ears will win your contest....that's not improving the signal up and above what the technology or the tools can allow but a subjective judgment call....

Well that's why I chose this sound, because it's a real challenge to make it sound good. What matters is to see how good it can be made. Probably not leaps and bounds better than what you heard, I know that, but it's all about pushing the limits of the new approach and seeing how it compares with all the more traditional denoising techniques.

[It'sJoeAgain]

Originally Posted by A_SN

Well that's why I chose this sound, because it's a real challenge to make it sound good. What matters is to see how good it can be made. Probably not leaps and bounds better than what you heard, I know that, but it's all about pushing the limits of the new approach and seeing how it compares with all the more traditional denoising techniques.

So far from what i've heard from the comparison samples and if in fact, u did it in only 5 minutes, then i give your software high marks for de-noising

[ncoak]

i'm curious as to whether anyone has successfully used photosounder for this type of cleanup, and what sort of results it yields? looks promising, although i'm not too impressed by the example on the website - was there ever a winner for the contest?

[finetuner]

Hey, thanks for digging up this thread, i'd been searching for it. Interesting stuff!

Originally Posted by ncoak

... was there ever a winner for the contest?

Yes, you can hear it here: Photosounder (scroll to the middle), including a GIMP recipe.
It's a bit cleaner again than the example in the contest by the O.P.

For a serious conservation/restoration request such processing may be far too deep, but considering the method, i think the result is very good.

[ncoak]

ah, thanks. interesting result, but not what i had hoped for...too many "space monkey" artifacts, as another user here calls them. also sounds like the flanginess you get as a reslut...er....result of absolute subtraction of the problem frequencies is pretty much identical to over-application of plugins like the waves x-noise. i actually find the original to be more intelligible than the 'cleaned up' version. cool program nonetheless.