Modelers and Aliasing

[BenIfin]

So that's it then, no room for further discussion?

Hey James ....... not at all ... was just reporting back my findings using Jays Sweep Test Method from his Post # 111 above and what he said I should listen out for in terms of aliasing :)

All the best,
Ben

 

[jay mitchell]

UPDATE

Well .... that didn't take long at all ... and was a lot easier to do and understand ..... did it nice and loud through my usual NS10's and Celestion / Matrix combo and headphones ... all 3 up very loud ..... had both in the same track so was able to A/B them instantly against each other back and forth multiple times.

It really matters very little what you listen through, nor is it necessary to listen particularly loud. If aliasing is present to an audible extent - which is what we're interested in knowing - it will be immediately obvious.

=> with both Helix Native -and- Tonex Native ..... to the best of my hearing .... there was absolutely no semblance of any kind of tones moving downward from a higher pitch to a lower pitch as the sweep wav file played all the way through ....even if it was somewhere deep in there, it was absolutely and completely in-audible to me.

That's the ticket.

So ...... hopefully my conclusion is right this time (?) when I say ..... neither Helix Native or Tonex produced any Aliasing of any "hear-ability" or consequence (?)

At least not with the amp blocks you used. Internally-generated aliasing - as opposed to aliasing in the A/D output - is caused by nonlinear processes, which add frequencies additional to and higher than the ones contained in the input. When a system produces this kind of aliasing, it will get worse as you increase the levels of distortion that the system generates.

That is not to say that other types of undesirable behaviors are absent, only that the undesirable behaviors don't include aliasing.

 

[Deadpan]

I have ran the test file with Nam and Tonex so far. Nam seems to have a higher rate of aliasing and often louder but both had the artifacts Jay described, but only at the higher frequencies.

 

[bigcupholder]

Uhh, no.

Wanna bet?

 

[James Freeman]

Here's a link to the sweep I use for this test: Alias Test.

Excellent test tone Jay.

 

[jay mitchell]

Wanna bet?

Sure. Wanna bet that AI can unscramble an egg while we're at it?

 

[Orvillain]

Sure. Wanna bet that AI can unscramble an egg while we're at it?

It could do it, but there'd be tiny little bits of egg shells around the upper limits of the system, and as much as most people would say they couldn't taste it, there would always be some super egg aficionado who would swear down that their analog eggs are way betterer coz they're fewer gritty.

 

[bigcupholder]

Sure. Wanna bet that AI can unscramble an egg while we're at it?

Have you ever worked with neural networks or have any idea how they work? I don't think it's absurd to suggest that all the extra feature channels could effectively learn to do the same thing as oversampling.

 

[pipelineaudio]

And don't forget the monkey wrench in the mix in the couple of famous cases where aliasing gave the product the "hardness" that sold it.

 

[Mongillo19]

monkey wrench

Foo Fighters used modelers confirmed

 

[AlbertA]

[2102.07757] How Convolutional Neural Networks Deal with Aliasing

In the second, we show that an image classifier CNN while, in principle, capable of implementing anti-aliasing filters, does not prevent aliasing from taking place in the intermediate layers.

Real-Time Guitar Amplifier Emulation with Deep Learning

This article investigates the use of deep neural networks for black-box modelling of audio distortion circuits, such as guitar amplifiers and distortion pedals. Both a feedforward network, based on the WaveNet model, and a recurrent neural network model are compared. To determine a suitable...

www.mdpi.com

It appears that even though the models were trained with non-aliased data, they suffer from aliasing, as was also
observed in [25]. At low and middle frequencies, the aliased components are 40 dB softer than the first
harmonic component, but at frequencies above about 15 kHz, some aliased components are louder than
their nearest non-aliased components. However, while the aliasing is evident with a high-frequency
sinusoidal input, no clear aliasing could be heard in the guitar and bass sounds processed through
the models. If this becomes a problem, it is possible that aliasing suppression techniques, such as
oversampling [ 40 ], could be applied during the neural network training, however, this is left for
future work.

Have you ever worked with neural networks or have any idea how they work? I don't think it's absurd to suggest that all the extra feature channels could effectively learn to do the same thing as oversampling.

That’s just a few samples.

 

[bigcupholder]

[2102.07757] How Convolutional Neural Networks Deal with Aliasing

Real-Time Guitar Amplifier Emulation with Deep Learning

This article investigates the use of deep neural networks for black-box modelling of audio distortion circuits, such as guitar amplifiers and distortion pedals. Both a feedforward network, based on the WaveNet model, and a recurrent neural network model are compared. To determine a suitable...

www.mdpi.com

That’s just a few samples.

Yeah I'm aware that aliasing occurs in neural networks. I mentioned in a previous post that there were papers showing that 2 years ago. My point was that it might be able to learn anti-aliasing with a proper objective function (i.e. an additional loss during training). Most tasks that neural nets are used for aren't sensitive to aliasing so there hasn't been much work done in this area.

Intuitively, the extra feature channels could be used by the network to do something similar to oversampling ("extra" in the sense that the neural net doesn't just have one signal inside it at intermediate layers). That would be similar to how we can do pooling by concatenation - the spatial resolution is subsampled but no information is lost because it's converted into extra feature channels. The network could learn to use the redundancy of the feature channels to effectively oversample if given a loss function that enforces that behavior. That's my guess anyways. I've worked in this field for a decade and published papers so I have a reasonable idea of what I'm talking about, but I could be wrong.

 

[jay mitchell]

"At low and middle frequencies, the aliased components are 40 dB softer than the first
harmonic component,"

First, that's inadequate. Second, what levels would the aliased components of the original processes have been when they were executed in the same sample space? It would be quite unremarkable for a modeler to produce aliased artifacts at no lower than -40dB. IOW, exactly how much an improvement in aliasing performance did the neural network produce?

"but at frequencies above about 15 kHz, some aliased components are louder than
their nearest non-aliased components."

Ayup. And saying that "no clear aliasing" could be heard in the sounds must be taken with a grain of salt.

 

[AlbertA]

Here's a comparison of the Aliasing performance between ToneX software vs ToneX Pedal vs Axe-Fx III using @jay mitchell's aliasing test signal.

The ToneX software plugin and pedal are running the JCM800 capture by SDS and the Axe-Fx III is running the JCM800 model matching the settings visible in SDS's video (which puts in the neighborhood gain and tone wise).

First ToneX Software plugin running at 48KHz, then ToneX Pedal, then Axe-Fx III


I don't hear any aliasing in the Axe-Fx III case.
The ToneX Pedal seems to be running at greater than 48KHz, since there's less aliasing (to my ears) than the software plugin running at 48KHz.

 

[BenIfin]

Here's a comparison of the Aliasing performance between ToneX software vs ToneX Pedal vs Axe-Fx III using @jay mitchell's aliasing test signal.

The ToneX software plugin and pedal are running the JCM800 capture by SDS and the Axe-Fx III is running the JCM800 model matching the settings visible in SDS's video (which puts in the neighborhood gain and tone wise).

First ToneX Software plugin running at 48KHz, then ToneX Pedal, then Axe-Fx III


I don't hear any aliasing in the Axe-Fx III case.
The ToneX Pedal seems to be running at greater than 48KHz, since there's less aliasing (to my ears) than the software plugin running at 48KHz.

Hey !

Man you must have some amazing hearing .... I played them back several times .... and chopped and changed etc.... and I could not hear any "aliasing descending tone" of any type or description as Jay pointed out to look for in his post.

Even if it is there at some miniscule level - which I cant hear - it is, as Jay put it, "inconsequential" :)

All the best and thanks for loading your test samples :)

Ben

 

[FractalAudio]

IMO a better way to hear aliasing is to sweep from 10-20 kHz. The aliasing tones will be below the sweep and you'll hear them move up and down in frequency.

 

[signalpath]

Hey !

Man you must have some amazing hearing .... I played them back several times .... and chopped and changed etc.... and I could not hear any "aliasing descending tone" of any type or description as Jay pointed out to look for in his post.

Even if it is there at some miniscule level - which I cant hear - it is, as Jay put it, "inconsequential" :)

All the best and thanks for loading your test samples :)

Ben

If you listen with headphones I’m pretty sure you’ll hear it. There’s a warbling pulse that descends in pitch while the test signal ascends in pitch, especially in the last third of the sweep.

 

[Boudoir Guitar]

I have ran the test file with Nam and Tonex so far. Nam seems to have a higher rate of aliasing and often louder but both had the artifacts Jay described, but only at the higher frequencies.

Yeah, that's easy audible. UFO warble like tone underneath the sweep.

 

[BenIfin]

If you listen with headphones I’m pretty sure you’ll hear it. There’s a warbling pulse that descends in pitch while the test signal ascends in pitch, especially in the last third of the sweep.

Well ..... still no :) I used my A/T monitoring headphones and the -only- difference I -can- hear between the Tonex Software and Tonex Pedal tones -vs- the Axe 3 Tones, is a "slight tonal difference" - but that is totally due to the fact that the Axe 3 Model used will clearly be different to whatever JCM 800 Model was Captured in the Captures.

I honestly cant imagine anyone picking or hearing the "aliasing differences" in these 3 tones if played to them, not knowing upfront which was which.

All the best,
Ben

 

[AlbertA]

Well ..... still no :) I used my A/T monitoring headphones and the -only- difference I -can- hear between the Tonex Software and Tonex Pedal tones -vs- the Axe 3 Tones, is a "slight tonal difference" - but that is totally due to the fact that the Axe 3 Model used will clearly be different to whatever JCM 800 Model was Captured in the Captures.

I honestly cant imagine anyone picking or hearing the "aliasing differences" in these 3 tones if played to them, not knowing upfront which was which.

All the best,
Ben

You don’t hear the warble undertones after like half the sweep? Like wwwwawwwawwww underneath the ascending sweep? Sounds kinda like Pac-Man.

They are low but they are there.