Modelers and Aliasing

[Foxmeister]

Still don't really know how to "read" which "spikes" are which .... ie:-

=> given I selected 10097k @ 0.00db ...... which are the "correct spikes" that are supposed to be there and which are the "aliasing spikes" which we dont want / want as low as possible ?

I'm glad I'm not alone in this!

 

[Enchilada Jones]

@James Freeman how would all this work with a real amp through a load and speaker sim?
Like if I play an amp through the Two Notes Captor X with speaker ir and power amp simulation and sent the line out into this software. Would you expect there would or could be an artifact problem or is the source being a real amp preventing that potential problem?

 

[FractalAudio]

^^^ You're not kidding ..... Voxengo SPAN has a hold function but it does not have a peak-hold ...... so "for fun" I "manually" very slowly swept the Full Freq Range in the Sine Generator Plugin back and forth multiple times from 20hz <-> 20k ...... the SPAN displayed "results" massively and wildly varied with even just very minor changes ...... I did notice - and I've no idea why - but at ~ 8k there seemed to be little to bugger-all aliasing showing up in SPAN, and what was there, was at "worst" -70db to much lower !?!? ..... this stuff does my head in :)

Ben

8K is a sub-multiple of the sampling frequency (assuming it's 48K). Sub-multiples of the sampling frequency place the aliasing products at zero, the frequency itself, harmonics or pi so they're not visible.

For example, if the frequency is 8kHz and the sample rate is 48kHz the 8kHz will produce harmonics at 8, 16, 24, 32, 40, 48, ... kHz. Assuming no oversampling (which it seems that these NN-based products don't use oversampling) then anything over 24kHz will alias.

32kHz will alias to 48-32 = 16kHz which ends up on the second harmonic. 40 kHz will alias to 48-40 = 8kHz which ends up on the fundamental. 48 kHz aliases to 0. Etc.

This is why it's important to sweep over a range. I use 10kHz - 20kHz because then anything below 10kHz is aliasing (or power supply stuff) and it's easy to measure.

 

[James Freeman]

Another small step towards understanding.
Regarding masking and audibility of quieter signal when there are louder signal along with it;

I've created a small listening test in my DAW and arrived to several understandings.
1. The more harmonically rich/dense the tone is, the better it masks quieter signals.
2. The audibility of quieter signals is better at higher SPLs but it depends on personal tolerance of the louder signal along with it.
3. My ears are most sensitive to sound between 1kHz and 6kHz.

This is probably the fundamentals of lossy audio compression but I think it directly applies to how audible aliasing is in the context of the harmonically rich signal of a distorted guitar amp.

Keep your hand on the volume control when listening to this.




With a pure sine tones I can hear and tolerate a second sine signal of -60dB, lower than that and the louder fundamental signal starts to hurt.
With a saw tooth signal I can't hear the second tone below -40dB at any volume.
Along with my previous post where I measured the peak spectrum of various typical guitar signals, now I know what to look and listen for.

Now for some test with plugins and tones... stay tuned.

 

[James Freeman]

There are frequencies that can yield almost no aliasing and increasing or decreasing the frequency by a tiny bit then yields large amounts of aliasing.

Yeah I've noticed that too.
For example 12kHz is an integer of 48kHz and Helix Native doesn't produce aliasing at that frequency at all.
But at 16kHz all hell breaks loose.

 

[James Freeman]

Two Notes Captor X with speaker ir and power amp simulation and sent the line out into this software. Would you expect there would or could be an artifact problem or is the source being a real amp preventing that potential problem?

If there is an IR Load there is a ADC/DAC and digital processing, yes this stage can introduce aliasing because you feed it harmonically rich signal.

Artifacts are only artifact if you can hear them or if they cause even bigger artifacts down the chain, which is what I'm trying to figure out in this thread.

 

[MirrorProfiles]

Getting ahead of myself here, but once this topic is done, I think testing the different kinds of oversampling algorithms would be quite interesting.

I think there’s definitely some sort of trade off, where you accept there will be some aliasing but hopefully at a point where it’s imperceptible. I think some anti aliasing filters can have more of a detrimental effect than just having a little bit more aliasing - they can mess with the phase/timing of frequencies which can have a result on transient detail. On some plugins, the phase effects of the anti aliasing filters can be a nightmare if you’re doing things in parallel. It’s another thing that different companies seem to implement in different ways.

 

[DrewJD82]

 

[deathbyguitar]

I have nothing to contribute to this conversation other than this kind of stuff is my favorite kind of thing to read.

 

[James Freeman]

The Daily TGF Science Journal got you covered!
So much nauseating gear science discussion you'll never want to visit other forums ever again!
I never leave town without my high voltage crystal oscillator, oscillograph and geiger counter just in case I have to measure a pedal.
Tell your friends!

 

[jellodog]

You're our resident expert on oversampling though

 

[northern_fox]

@James Freeman have you tried this with a lower ESR NAM capture? 0.02x (from the other thread) is still pretty high. I'd love to see this with a more accurate model - for example, this recto red capture with an ESR of 0.0027 (Data from Jon Arnold and Petr Canov - I trained this test version)

(Also, maybe we can allow *.nam uploads here ???)

 

[James Freeman]

Hi @northern_fox.
Lower ESR is not related to aliasing.
Currently all neural network based captures display high aliasing, it's a fundamental trait of the technology, particularly the capture player.
Maybe they (Neural DSP, ToneX, NAM) can fix this by oversampling and applying a nyquist filter but AFAIK none of them do.

PS.
We have a NAM thread: https://thegearforum.com/threads/nam-neural-amp-modeler.1698/
I will try your recto capture.

 

[northern_fox]

Hi @northern_fox.
Lower ESR is not related to aliasing.
Currently all neural network based captures display high aliasing, it's a fundamental trait of the technology, particularly the capture player.
Maybe they (Neural DSP, ToneX, NAM) can fix this by oversampling and applying a nyquist filter but AFAIK none of them do.

PS.
We have a NAM thread: https://thegearforum.com/threads/nam-neural-amp-modeler.1698/
I will try your recto capture.

I don't have a good understanding of where the aliasing comes from in the processing chain (I know how it works fundamentally) - I kind of just assumed a more accurate model would exhibit less of it. Appreciate your explanation

 

[la szum]

"Aliasing.... aliasing.... I hear aliasing..... aliasing over my head."

 

[JiveTurkey]

 

[bigcupholder]

Oh, and I should add that every neural network based product I've tested has tremendously bad aliasing performance.

Hi @northern_fox.
Lower ESR is not related to aliasing.
Currently all neural network based captures display high aliasing, it's a fundamental trait of the technology, particularly the capture player.
Maybe they (Neural DSP, ToneX, NAM) can fix this by oversampling and applying a nyquist filter but AFAIK none of them do.

PS.
We have a NAM thread: https://thegearforum.com/threads/nam-neural-amp-modeler.1698/
I will try your recto capture.

Aliasing isn't a fundamental trait of neural networks, it's just something that ML researchers weren't aware of at all for most applications. There were a couple computer vision papers at NeurIPS 2021 (IIRC) that showed there's tons of aliasing in the feature maps of convolutional networks after max-pooling or strided convolutions... but it didn't really affect results for classification tasks significantly and I haven't seen any follow up on it. Relatively very little ML research is done on generating audio, so maybe anti-aliasing is well known in that community, but most ML people don't have a signal processing background.

You can apply a window function to the filters, basically baking in an anti-aliasing filter. It doesn't even add any extra compute once the model is trained. I don't see why you'd need to oversample for a neural net but maybe I'm missing something there.

 

[paisleywookiee]

I don't know wtf is going on, but I'm decidedly anti-alias. Down with the alias!

 

[fretworn]

 

[pipelineaudio]

Some idiot made this one a few years ago