I'm looking for something with a slightly better saturation character than tanh~, but with not much more than about 4 times the CPU usage. Was thinking of just upsampling and filtering tanh~ in a subpatch, but does anyone else have a decent saturation algorithm that might be better?
hello, I think "nicer" is really subjective. tabread4~ is the faster algorithm. (then, put whaterver curve you want in the table)
you can have a look at nusmuk-audio for different distortion algorithm, cheers cyrille
Le 27/10/2017 à 20:37, Matt Davey a écrit :
I'm looking for something with a slightly better saturation character than tanh~, but with not much more than about 4 times the CPU usage. Was thinking of just upsampling and filtering tanh~ in a subpatch, but does anyone else have a decent saturation algorithm that might be better?
Pd-list@lists.iem.at mailing list UNSUBSCRIBE and account-management -> https://lists.puredata.info/listinfo/pd-list
ok yeah, i guess i mean, "nicer" in a way that doesn't alias so bad.
seems oversampling and filtering is the way to go with that, but just wondering if there's some cheap way to avoid aliasing with distortion, without oversampling too much?
depending on the input you can also LP filter the signal before the distortion.
oversampling and filtering is the worst way to deal with aliasing, but sometimes there is no other solution.
Le 27/10/2017 à 22:03, Matt Davey a écrit :
ok yeah, i guess i mean, "nicer" in a way that doesn't alias so bad.
seems oversampling and filtering is the way to go with that, but just wondering if there's some cheap way to avoid aliasing with distortion, without oversampling too much?
Also try arctangent in place of tanh. Its asymptotic behavior is 1/x instead of e^(-x) which makes for softer clipping.
I often make a big table to speed up the atan lookup using tabread4~ although you have to be careful at the table limits.
cheers Miller
On Fri, Oct 27, 2017 at 11:12:56PM +0200, cyrille henry wrote:
depending on the input you can also LP filter the signal before the distortion.
oversampling and filtering is the worst way to deal with aliasing, but sometimes there is no other solution.
Le 27/10/2017 à 22:03, Matt Davey a écrit :
ok yeah, i guess i mean, "nicer" in a way that doesn't alias so bad.
seems oversampling and filtering is the way to go with that, but just wondering if there's some cheap way to avoid aliasing with distortion, without oversampling too much?
Pd-list@lists.iem.at mailing list UNSUBSCRIBE and account-management -> https://lists.puredata.info/listinfo/pd-list
2017-10-27 19:12 GMT-02:00 cyrille henry ch@chnry.net:
depending on the input you can also LP filter the signal before the distortion.
oversampling and filtering is the worst way to deal with aliasing, but sometimes there is no other solution.
recently I found out the results were kinda terrible and there wasn't much I could do about it, oversampling it to an insane amount like 512x didn't improve significantly at all...
so I was wondering: - 1) why? Is there any other way?
You say sometimes "no", but when can you? I assume you could do it better if you're designing oscillators and things like that, but for this kind of alias generated by digital audio manipulations, maybe there's nothing else at all? is that it?
thanks cheers
Le 28/10/2017 à 02:08, Alexandre Torres Porres a écrit :
2017-10-27 19:12 GMT-02:00 cyrille henry <ch@chnry.net mailto:ch@chnry.net>:
depending on the input you can also LP filter the signal before the distortion. oversampling and filtering is the worst way to deal with aliasing, but sometimes there is no other solution.recently I found out the results were kinda terrible and there wasn't much I could do about it, oversampling it to an insane amount like 512x didn't improve significantly at all...
so I was wondering: - 1) why?
if you start with a 1KHz sinuzoid, and add 10 harmonics, the last harmonics frequency is 1K * 2^10, so about 1MHz and this is just a "small" distortion.... according to what I read on the web, a good oversampling factor is between 2^8 and 2^14.
So, a 512x oversampling is not perfect but should improve a lot. (I suspect a patch problem if you did not have any improvement)
Is there any other way?
if you don't want to oversample, you must manipulate the sound in a way that did not generate harmonics higher that SR/2 this is what antialiased oscillators are doing.
You say sometimes "no", but when can you? I assume you could do it better if you're designing oscillators and things like that, but for this kind of alias generated by digital audio manipulations, maybe there's nothing else at all? is that it?
I did not find any interesting algorithm searching on the web, but I may have missed something.
cheers c
thanks cheers
Pd-list@lists.iem.at mailing list UNSUBSCRIBE and account-management -> https://lists.puredata.info/listinfo/pd-list
On Sam, 2017-10-28 at 10:48 +0200, cyrille henry wrote:
Le 28/10/2017 à 02:08, Alexandre Torres Porres a écrit :
2017-10-27 19:12 GMT-02:00 cyrille henry <ch@chnry.net mailto:ch@c hnry.net>:
depending on the input you can also LP filter the signal before the distortion.
oversampling and filtering is the worst way to deal with aliasing, but sometimes there is no other solution.
recently I found out the results were kinda terrible and there wasn't much I could do about it, oversampling it to an insane amount like 512x didn't improve significantly at all...
so I was wondering: - 1) why?
if you start with a 1KHz sinuzoid, and add 10 harmonics, the last harmonics frequency is 1K * 2^10, so about 1MHz
Hm... the tenth harmonic of the fundamental frequency 1kHz is actually 11kHz, not 1MHz, or am I missing something here? 2kHz is the first partial, 3kHz the second, etc.
Roman
Le 28/10/2017 à 11:31, Roman Haefeli a écrit :
On Sam, 2017-10-28 at 10:48 +0200, cyrille henry wrote:
Le 28/10/2017 à 02:08, Alexandre Torres Porres a écrit :
2017-10-27 19:12 GMT-02:00 cyrille henry <ch@chnry.net mailto:ch@c hnry.net>:
depending on the input you can also LP filter the signal before the distortion.
oversampling and filtering is the worst way to deal with aliasing, but sometimes there is no other solution.
recently I found out the results were kinda terrible and there wasn't much I could do about it, oversampling it to an insane amount like 512x didn't improve significantly at all...
so I was wondering: - 1) why?
if you start with a 1KHz sinuzoid, and add 10 harmonics, the last harmonics frequency is 1K * 2^10, so about 1MHz
Hm... the tenth harmonic of the fundamental frequency 1kHz is actually 11kHz, not 1MHz, or am I missing something here? 2kHz is the first partial, 3kHz the second, etc.
I think added harmonics are : 2, 4, 8, 16...
Roman
Pd-list@lists.iem.at mailing list UNSUBSCRIBE and account-management -> https://lists.puredata.info/listinfo/pd-list
2017-10-28 9:51 GMT-02:00 cyrille henry ch@chnry.net:
I think added harmonics are : 2, 4, 8, 16...
No, they're all multiples of the fundamental frequency. You can debate wether the 1st harmonic or 1st partial is the fundamental, people use this in different ways. I see the fundamental as the first partial component and the first harmonic on top op that, but usually just consider both as the fundamental, so 10th harmonic would be 10Khz (or 11Khz).
cheers
Le 28/10/2017 à 14:27, Alexandre Torres Porres a écrit :
2017-10-28 9:51 GMT-02:00 cyrille henry <ch@chnry.net mailto:ch@chnry.net>:
I think added harmonics are : 2, 4, 8, 16...No, they're all multiples of the fundamental frequency. You can debate wether the 1st harmonic or 1st partial is the fundamental, people use this in different ways. I see the fundamental as the first partial component and the first harmonic on top op that, but usually just consider both as the fundamental, so 10th harmonic would be 10Khz (or 11Khz).
yes, sorry, I made a mistake. if your distortion is symmetrical, it can't produce even harmonics, so added harmonics are 3, 5, 7, 9... see patch or snapshot : the 4th partial of 350Hz is 1500Hz and is not present in the distorted sinus.
(if distortion is not symmetrical, then you are right and produce all partial)
So, oversampling by 512 must kill all aliasing.
cheers c
cheers
Pd-list@lists.iem.at mailing list UNSUBSCRIBE and account-management -> https://lists.puredata.info/listinfo/pd-list
2017-10-28 11:12 GMT-02:00 cyrille henry ch@chnry.net:
So, oversampling by 512 must kill all aliasing.
yeah, but what I found is that it doesn't get that much better than, say 16x
I don't understand. can you share your patch? thanks c
Le 28/10/2017 à 16:54, Alexandre Torres Porres a écrit :
2017-10-28 11:12 GMT-02:00 cyrille henry <ch@chnry.net mailto:ch@chnry.net>:
So, oversampling by 512 must kill all aliasing.yeah, but what I found is that it doesn't get that much better than, say 16x
The trick is having a very good anti-aliasing filter just like the FIR in a high end ADC or SRC or very good mastering processors. An IIR just won't work well due to the phase shifting and low order stop band attenuation. I don't think Pd has anything built in other than fexpr~ and coefficient selection is something of a black art to trade off between the attenuation, taps/delay, transition band, impulse response etc.
On Fri, Oct 27, 2017 at 8:08 PM, Alexandre Torres Porres porres@gmail.com wrote:
2017-10-27 19:12 GMT-02:00 cyrille henry ch@chnry.net:
depending on the input you can also LP filter the signal before the distortion.
oversampling and filtering is the worst way to deal with aliasing, but sometimes there is no other solution.
recently I found out the results were kinda terrible and there wasn't much I could do about it, oversampling it to an insane amount like 512x didn't improve significantly at all...
so I was wondering: - 1) why? Is there any other way?
You say sometimes "no", but when can you? I assume you could do it better if you're designing oscillators and things like that, but for this kind of alias generated by digital audio manipulations, maybe there's nothing else at all? is that it?
thanks cheers
Pd-list@lists.iem.at mailing list UNSUBSCRIBE and account-management -> https://lists.puredata.info/ listinfo/pd-list
-
Alexandre Torres Porres -
Chris Clepper -
cyrille henry -
Matt Davey -
Miller Puckette -
Roman Haefeli