Cuki's Modeling/Convolution Experiments

[Mischief]

Originally Posted by Cuki79

Hi everyone,



Now the process enables to tune separately 3 bands. On each band you can tune the length of the notes and its amplitude. I don't put anymore the parameters (they are 11)



Trance Amulet alone (mono):





External microphones (stereo):





3bands IR:





The new spectrum IR is like that:





If you remember from previous post with old rectification process it basically looked like that (maybe a bit improved in the last posts)





So the 3 band process enables me to clean the midrange without compromizing the low end. In the last posts, I only cleaned the high end.



I've discovered that too much cleaning in the midrange is not good. Here there are still a lot of small "accidents" in the spectrum but keeping these midrange "ringing" gives a bit of metallic reverb that microphone takes usually have... Well I tried to keep that. (But I can make it sound clinical and very straight now if I want). I also gave the midrange a slight boost, because it was lacking.



The high end gets an extra cleaning with a very short IR/reverb tail and a firm boost to compensate for.



Now the sum of the 3 bands (green) is very smooth compared to the previous IR spectrum. You can't really see it at the intersection of the bands, but it is much smoother.



Now improving the sound is mostly about tuning the 11 parameters. (amplitude, start time of decay, decay rate x 3 bands + 2 crossover frequencies)



If SpruceTop is still reading the thread I'd be to glad give a try to the Taylor ES2. I have a ES1.3 on my Taylor and if it sounds good I may pay the 300 bucks for upgrading to ES2 (It can't work on magnetic systems). You only have to send me 4 wav files.



Cuki

You are doing great things Cuki,
I have been inundated with time sensitive projects. Whenever I have had a tiny bit of free time I've had to put to getting my recording gear organized and mounted so I can do some recording again. Hopefully that will be soon and I can send you some fresh samples.

I'm always amazed at how you have been able to improve your technique and the results at each turn.

[SpruceTop]

Originally Posted by Cuki79

If SpruceTop is still reading the thread I'd be to glad give a try to the Taylor ES2. I have a ES1.3 on my Taylor and if it sounds good I may pay the 300 bucks for upgrading to ES2 (It can't work on magnetic systems). You only have to send me 4 wav files.

Cuki

Hi Cuki,

I'll see what I can come up with for you. You just need wave files created from the output of the ES2 (and, later, other systems), and not any microphone tracks, correct?

Ken

[Cuki79]

Hi Sprucetop

Thanks for the help, I've sent you a private message with my personal address for the google drive and the detailed procedure for the wav files.

Thanks again,
Cuki

[Doug Young]

If any one is still following this, more experiments as they happen! Here's Cuki's latest, an IR created from a mono sample of the Trance and the mic'd guitar. I tried this IR with the IPSi pedal and Sound Designer in Logic, both of which sounded very similar. Here's a sample:

Dry mono Trance recording:



Then applying the IR to the same sample using Sound Designer:



Cuki, this is your "3Bands_10min3i_TOP" IR

[Mischief]

I am still following Doug,
What are your current thoughts so far?

Cuki seems to be able to tweak things better and better all the time.

[Doug Young]

Yeah, things keep improving, so I remain intrigued :-) I am starting to hear some of the same resonances and weird artifacts that I have always found objectionable in the Aura, tho. I'm a bit torn. On one hand, raw pickups never sound quite natural. At the same time, in a live setting, a direct dry sound often seems to work better than any processing we can do. I feel like so far, these IRs can transform a pickup recording to be competitive with a bad mic recording, which is an improvement :-) But not all the way there yet to a good mic recording, and in a live performance might be a step backward. But it's interesting to try and see what's possible.

[Cuki79]

Hi everyone,

here are my latest IRs.

Since Doug said the last IRs matched poor microphone take, I started to listen carefully and it is very true the last IR was kind of caricatural of mic sounding. The biggest problem is the midrange.



So instead of using 3 bands, I am now using a near continuum EQ (I don't remember the number of points but it's much more but not too much. The EQ curve (in red) is based on a 2nd turn of IR matching made from a single note.

As you can see the EQuing tends to compensate for the overexagerate cuts that the IR does at 2 kHz and 4 KHz. (Not that spectrum smoothing is done by cut later on in the process). The result is closer to a Stereo Amulet than a cheap Mic take.

Moreover, Doug told me that there were too many resonances (artifacts) that ring when he was playing up the neck. So I went chasing for those. He made a special recording to let me actually spot them.



Here you can see on the left of the picture above many sharp peaks going very high. Those tends to feedback and ring... So I Equed them out. Note there are still many in the upper frequencies but we don't hear them much... But in live situation, they could probably cause enormous feedback... This has to be tested LIVE to spot them all.

Enjoy the listening (Doug's playing)

The Trance Amulet alone:


The external microphone take:


IR shot 1, EQ, cleaned and then recut.


IR shot 2, EQ (red curve previously presented), cleaned and then recut.


I've read very interesting things lately so I may try a total new process based on things scientists do for image deconvolution.... However it may take a little time to get that to work.

Updated:
Sorry some filter was engaged... But it was probably also when I did the second IR round... I am trying to use space designer as Doug so we hear the same thing... But there buttons everywhere...

IR shot 1:


IR shot 2:


Well it could be improved, if I start the whole process again.
Cuki

[MaurysMusic]

This is really interesting - thanks again for "boldly going where no man has gone before"!

[Doug Young]

My apologies for all the noodling :-)

[Cuki79]

Just to give hints of what I want to try: When you deconvolve, the high end is dominated by noise. It is due to the fact that we have very few signal there. So the noise from the microphone recording will be devided by a very small number for example 0.001. If you divide by 0.001, it's like multiplying by 1000. As the highest note we play is around 1 kHz, there are not much energy over 5 kHz (even the harmonics coul'd be few dBs above 0). As a result, the deconvolution process amplifies the noise greatly.

People from imaging have the eaxct same problem when debluring images. They want to get back sharp edges, which is for us the high end of the frequency spectrum.

In this example, the two top images are blured. I think in this case they did not have the two images (blind deconvolution) meaning that they only had the piezo signal.

The 2nd row use an algorythm that exploits geometric things (I guess) and the last one use deconvolution with some fine mathematic strategy. The contrast (high frequencies) are partially recovered.

Another example here (see figure below) where they had both the blurred and blurring process but try to restore the original image. For us it would mean they had the "pickup" signal and "impulse response" and try to recover the "mic" signal with the high end. First image on top left would be the pickup, third image on the right would be the image +IR with the process I want to try.



Maury is right, no man did it since a Turkish girl doing her Phd in Lausanne (Switzerland) made a 1-D sound compatible version of that.... But she never tried to address our problem, probably she never played guitar with a piezo pickup... People usually uses 1-Dprocess for seismic detection for example, something more valuable than acoustic guitar modelling

Anyway it explains why the Aura image are so "undefined" and "blurry", and why you need to blend to restore the high end.

So there is still a long way before our quest ends...
It's going to take me more than one day, there is a bunch of math I don't really understand yet in that process. Possibly the one that ToneDexter uses.

PS: It's not going to be easy: An image has usually 256 level of dynamic (8 bits) we use over 16 millions level of dynamic (24 bits). For us: small is very very small and blurry might be very very blurry compared to images. But we can get information from time, they can't... Don't expect that it works next week.

Cuki

[Cuki79]

OK

Here is the new algorithm. No more Eq, except a small shelf filter under 110Hz that is not compulsory. 3 bands fine cut is not compulsory anymore (hear yourself). The raw IR (handcut) is pretty nice already and you don't need to kill the ringings anymore. (The new algorithm does not create such artefacts: the IRs passed the Doug Young's recording test at first try with no ringing where the former algorithm needed 5-7 notch filters)

Trance alone:


External Microphones:


New Algorithm space-designer hand cut: (please LOWER/ADJUST the VOLUME to match the mic recordings)


New Algorithm 3 bands cut: (please LOWER/ADJUST the VOLUME to match the mic recordings)


Weird is that the new algorithm totally failed with the lyric (I've just tried), in this case, the last IRs I've published are way better (may be it's todays recordings... since I changed mics and audio interface... to be continued)

I am happy with these IRs (Doug's)... and super frustrated that the process does not work with my guitar

Nota Bene: The new algorithm is not based on the image processing algorithm I've described in the last post... The one for image did work... but it did not do what I wanted mainly because our SNR is much worst than in images. ... Well it's a long story.... But I've learned tons of things doing it, and it led to this new algorithm that has nearly no parameters and works nicely with Doug Young's recordings.

Don't hesitate to comment.

[Cuki79]

Hi everyone,

since I have no news from Doug or SpruceTop, I thought I'd give some explanation on the parameters so everyone could get how the new algorithm affects the sound. Basically there are only two parameters they can be tuned differently with respect of frequencies (usually in my last posts Bass/Medium/Treble) but in the following I used the same for every frequencies so all the spectrum is treated the same way.

The first parameter is called M. For lower values you get a lot of "metallic resonances", the higher the values the less you have the metallic brightness.

The second parameter is the tail length, it's a parameter that can be tuned in most IR reverb plugin (like Space designer for example). The only difference is that usually I tuned it differently with respect to frequencies. Long reverb tails, give some kind of artificial room reverb. Note that combined with the metallic resonances it sounds horrible.

At the beginning, I've done few strokes so you can hear the reverb tail (it's horrible for low values: sorry).

M-parameter values: 50, 200, 500, 2000 and 20000 for long tail. You'll hear it gets better with higher numbers. Note that the last sound is the Lyric without IR (it's mono). (Sorry for the bad Maury imitation)


M-parameters for short tail.


The goal is to get the sweet spot, where you still get some metallic attack for definition. Find the correct reverb tail length to have a sound dry enough for strumming but keep some airiness for fingerpicking. (Of course tuning those parameters differently with respect of frequency band helps).

Finally I think the easiest way to fine tune, is to put all these IRs on the pedal and test. You quickly realize if you like the reverb tail or not and how much the metallic ringing bothers you (especially up the neck).

Now I am going to work on the tailing process, it deaden the sound too much for my taste.

Please if someone is still reading this thread just post something because I really feel I am doing all this for no one...

[MaurysMusic]

I'm following along but I can't guarantee that I'm fully understanding.

[Cuki79]

Thanks Maury

Doug will soon give me some feedback on the new process.

Another demo:

DPA 4099 (stereo spaced pair), Lyric alone (mono), New process IR (M-parameter 521, simply cut short), same with blend (about 50%).



There is a bit too much bass on the IR but it is an easy fix. The midrange is more problematic... did not find yet anything satisfying.

[Mischief]

Hi Cuki,
I am still here. Things have been completely hectic. I did pick up my Takamine guitar yesterday that had its refret job completed but I have not had a chance to play it other then to check it at the luthier.

I'm still fascinated in the progress here and will try and get some decent recording samples to you if you are still interested in applying some of these IR to the palathetic pickup to see how they respond.