Modelers and Aliasing

[pipelineaudio]

In b4 someone says things can’t alias if they sum.

 

[DrewJD82]

You're our resident expert on oversampling though



View attachment 5314

I’m totally fine taking on that moniker.

 

[imperator]

I'm not sure aliasing is important in this situation. NAM learns a map from a input buffer to the system output and minimizes, I'm guessing, the 2 norm so if it accomplishes this goal and there is some aliasing it almost irrelevant outside of some edge cases. If you really wanted we could maybe change the cost function to add a penalty, or add some training data where a tone is injected >fs/2

 

[FractalAudio]

I'm not sure aliasing is important in this situation. NAM learns a map from a input buffer to the system output and minimizes, I'm guessing, the 2 norm so if it accomplishes this goal and there is some aliasing it almost irrelevant outside of some edge cases. If you really wanted we could maybe change the cost function to add a penalty, or add some training data where a tone is injected >fs/2

Aliasing in important in every situation. Aliasing is the creation of harmonically unrelated and undesirable tones in the audible spectrum.

Aliasing is most easily heard when playing single notes. It is masked when playing chords but raises the effective noise floor causing a loss of clarity. IME it also causes rapid ear fatigue.

A real amp doesn't alias and good modelers have minimal aliasing. The most common way to reduce aliasing is to increase the sample rate either natively or by oversampling. You can also reduce aliasing using antiderivatives but this only works in very specialized cases when using waveshapers.

 

[Cirrus]

For my thinking, the best strategy is to use a high sample rate, AND be aware that nothing over 20k can be heard anyway, so be aware of where ultrasonic harmonics might be generated and roll them off before the next distortion stage that will cause the harmonics of those harmonics to appear - basically a big no-man's land between 20kHz and the Nyquist frequency that gets occasionally cleared out if you have concerns, with plenty of space to use less sharp filters for less phase shift/ ringing.

That's if you're worried about it, anyway. It often doesn't matter, and sometimes the sonic effect of avoiding aliasing will sound worse than just letting it happen.

Real amps do produce enharmonic information - wolf tones from your notes basically ring modulating against the 50/60hz mains frequency that is mostly, but not completely, removed by the filter caps, choke etc, and beat frequencies of multiple notes in chords/ double stops interacting. Somehow, that seems more musical than aliasing but just a little perspective that the analogue world is also full of little things like this.

 

[DLC86]

When I received my FM9 some months ago, I thoroughly compared it to my axe fx 3 before selling the latter, and one of the comparisons I made was about the aliasing performance. I haven't posted the results till now on other forums to avoid starting a flame war among both fanbois and haters, since these findings somewhat contradict what fractal has officially (un)said about this, and I perfectly understand that they avoid disclosing these details for those same reasons, on Fractal forum and TOP this would be blown out of proportion for sure.
But since we are on TGF and this interesting thread came out, here we go:

I used the method proposed by Cliff to measure the aliasing, feeding a 9-11 kHz sine sweep to the same exact amp model (VH4) and corresponding FWs between the two units. I also tested if there was a difference between having just 1 amp block vs 2 on the axe fx and it turns out there is indeed.

Axe FX with one amp loaded

Axe FX with two amps loaded (only one measured)

FM9

It's pretty clear and expectable that the one having less aliasing is the Axe FX with just one block loaded in the preset, it probably uses the whole 1 GHz horsepower of one of its core to process the amp alone, with an overkill amount of oversampling to reduce aliasing at minimum.
When loading two amp blocks that horsepower has to be shared between those two, so the oversampling rate is probably cut in half and the aliasing gets louder.
The FM9, having basically half the processing power of the Axe FX but split in 4 cores instead of 2, has to cut the oversampling in half again so it has the highest amount of aliasing among the three. And for the reason that the two amp blocks each have a dedicated core, running one or two amps in a preset doesn't make a difference.

The aliasing of the FM9 might seem quite high from these graphs since it's "just" 36 dB lower than the fundamental frequency, but it isn't really, cuz this test was made with ridiculously high amounts of gain to make the aliasing more visible in the RTA graph, something that nobody would likely ever use in real world (VH4 ch.3 with all the knobs cranked), and the dry sine sweep was at -12 dBFS, a level that probably no guitar DI signal can reach at those frequencies.
Making this test with more "real-world" settings and signals would push the aliasing below -60dB even on the FM9, and in fact I could not hear any difference at all between the two devices in my presets, that's why I decided to keep the fm9 and sell the Axe FX in the end, they really sound the same.

 

[DLC86]

Today I also compared the FM9 with NAM and Tonex, this time using the settings I normally use on the VH4 so the aliasing on the FM9 looks definitely better, but the sine sweep was still at -12 dBFS so that might still be a bit unrealistic, but for sure it showcases the huge difference between a modeler and a current NN capture.

FM9

NAM

Tonex

 

[Mongillo19]

Today I also compared the FM9 with NAM and Tonex, this time using the settings I normally use on the VH4 so the aliasing on the FM9 looks definitely better, but the sine sweep was still at -12 dBFS so that might still be a bit unrealistic, but for sure it showcases the huge difference between a modeler and a current NN capture.

FM9
View attachment 5359


NAM
View attachment 5358


Tonex
View attachment 5357

Would you need a graph of an actual VH4 or does what I said make 0 sense lolol

 

[James Freeman]

Thanks for sharing @DLC86.
Yeah -12dBFS is an unrealistic input value at 10k-20k range.
Also, according to my masking hearing test I can't hear any signal -40db below fundamental of a harmonically rich tone (distorted amp) so that is also something to consider.

EMG 81 produces like -85dB at the 10k-20k region, and drive pedals around -35dB when normalized.

 

[DLC86]

Would you need a graph of an actual VH4 or does what I said make 0 sense lolol

Well, real amps don't produce aliasing, so not for this purpose

 

[Mongillo19]

Well, real amps don't produce aliasing, so not for this purpose

Gotcha! I did know that atleast! I just wasn't sure if you'd need a graph of the VH4 to compare the other devices at to see where the differences might be

 

[James Freeman]

Gotcha! I did know that at least!

You would if you read the second post in this thread.
Plenty of good info here.

 

[DLC86]

Gotcha! I did know that atleast! I just wasn't sure if you'd need a graph of the VH4 to compare the other devices at to see where the differences might be

Well it might be interesting to compare a real amp to a capture of that actual amp, since these were captures of the fm9 instead it would be pointless cuz there could be lots of differences starting from knobs settings and tapers.

 

[Mongillo19]

You would if you read the second post in this thread.
Plenty of good info here.

You right you right. It's what I get for jumping in mid thread

 

[DLC86]

EMG 81 produces like -85dB at the 10k-20k region, and drive pedals around -35dB when normalized.

I'll repeat my second comparison with more realistic levels soon

 

[James Freeman]

BTW @DLC86, the sweep should be 10k-20k with a quick fade-in/out at the ends to prevent a 'pop' in the measurement.

Use this sweep I made:

https://drive.google.com/file/d/1qM5OrM40K-RfQfm0K3mbPvl-VGMrPmqc

1Min 10kHz-20kHz Sine Sweep 48kHz, Mono and normalized to -0.1dB with Fade In/Out.
Lower its level in DAW.

My span settings, I click the span window to reset it before hitting play in DAW.
Offset Normalized so I can see the aliasing -dB value on the db scale on the right.

Big difference in aliasing with realistic vs unrealistic input dB values at the 10k-20k range.

 

[DLC86]

BTW @DLC86, the sweep should be 10k-20k with a quick fade-in/out at the ends to prevent a 'pop' in the measurement.

Use this sweep I made:

https://drive.google.com/file/d/1qM5OrM40K-RfQfm0K3mbPvl-VGMrPmqc

1Min 10kHz-20kHz Sine Sweep 48kHz, Mono and normalized to -0.1dB with Fade In/Out.
Lower its level in DAW.

My span settings, I click the span window to reset it before hitting play in DAW.
Offset Normalized so I can see the aliasing -dB value on the db scale on the right.
View attachment 5360

Great! I'll use your file and your span settings (except the slope maybe) from now on, thanks!

 

[DLC86]

PS: it could be useful to sweep only on a limited range of frequencies and using a larger view on span though, cuz you can also see the first harmonics this way and also match the gain with those if you're comparing two different devices/plugins

 

[James Freeman]

Yeah... not looking good for NAM.

 

[imperator]

Yeah... not looking good for NAM.

View attachment 5362

Try running a low pass before the input to NAM, and compare the accuracy with and without the low pass. It's probably obvious that the accuracy will degrade, my point is that the aliasing may be combining in a way that minimizes the error in the case of the neural net.