Modelers and Aliasing

[James Freeman]

Try running a low pass before the input to NAM

It's a sine sweep from 10k to 20k, a sharp moving frequency peak, there is nothing to low pass.

 

[imperator]

It's a sine sweep from 10k to 20k, a sharp moving frequency peak, there is nothing to low pass.

No, the actual guitar signal.

My point is that in this context, youre taking a new input to the model that is outside of the distribution it was trained on and checking how it has generalized. That's a big nono in ML.

I still like the experiment though, it's fun.

 

[yeky83]

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.

How do you not add extra compute if you want to bake in an anti-aliasing filter... wouldn't it have to oversample to do so?

 

[imperator]

How do you not add extra compute if you want to bake in an anti-aliasing filter... wouldn't it have to oversample to do so?

You can combine the two filters so it's just one convolution

 

[yeky83]

Progress in small steps.

Spoiler: Pedals

View attachment 5277

Strumming the open strings with EMG 81 and capturing the spectrum of Bypass, Rat, and a Tubescreamer, all normalized to peak.
Pedals are set to everything at maximum to generate highest harmonic distortion above 10kHz, obviously real pedals don't have aliasing so we are seeing the peak level of pure analog harmonic distortion.

View attachment 5278


The numbers below are for the 10kHz to 20kHz range you can see on the graph above, the input frequency range that generates aliasing if processed by a modeler or a plugin.

EMG 81: -70dB to -84dB.
Tubescreamer: -52dB to -66dB.
Rat: -39dB to -48dB.

Reminding that these numbers are form a normalized signal, ie. if the Rat peaks at -10dB, the values would be -49dB to -58dB, moved down by 10.

Alright, so now we can see that running a huge test signal above 10kHz into the modeler or plugin to test aliasing is meaningless.
We need to use real world 10kHz+ values, whether it is a pickup, a boost pedal, or anything we put before the amp.

More testing coming.

How about... the Rat into a Dallas Rangemaster, bringing 10-20 kHz back to 0dB??

 

[yeky83]

You can combine the two filters so it's just one convolution

Maybe I'm the dummy here, but I'm not sure if you understand how anti-aliasing filters work...? You don't just filter it out, you first increase the sampling rate then filter it out. Otherwise it wouldn't incur any compute cost with traditional amp modelers either.

 

[imperator]

Maybe I'm the dummy here, but I'm not sure if you understand how anti-aliasing filters work...?

Oh boy, let me ask you this, how would you down sample a signal?

 

[yeky83]

Oh boy, let me ask you this, how would you down sample a signal?

Not sure what you're getting at... are neural networks inherently oversampled or something? Added more to the above post after you replied.

 

[signalpath]

IME it also causes rapid ear fatigue.

I don’t know if it’s down to aliasing but I experience a seriously debilitating effect when using modellers over prolonged period of time. Specifically when using headphones and FRFRs. It was really pronounced with the HX family prior to the oversampling update but it remains a factor.

I’m interested in the aliasing comparison of Axe-FX III to the FM9 and wonder how they compare to the FM3 which is what I have. In my experience the FM3 is better than the HX but I still experience residue effects over time.

The cumulative effect of using a modeller with headphones and FRFRs is that my attention gets scrambled and actually takes a couple of days to return to normal. I know that this isn’t a common experience and I’m certainly not making any claims about the acuity of my hearing because I’m certain many others have better ears than I do. But it is most definitely a neurological response. The issue came up on TP when I was a member there. There was one other person who reported a similar experience.

At which point I was on the verge of giving up on modelling but then tried running my HX Stomp into the effects loop of an amp with no IRs. And the disorientation disappeared. I subsequently tried the Stomp with a PS-170 and a regular guitar cab and again experienced no adverse effects.

I mention all this because I wonder if it’s indeed aliasing that’s behind what I’m experiencing, why is that running the modeller through a regular cab corrects the problem. And also, is it the reason that people who may not suffer the extreme reaction I have but are subtly experiencing the same effect have a marked preference for using a modeller through a regular guitar cab.

 

[DLC86]

No, the actual guitar signal.

My point is that in this context, youre taking a new input to the model that is outside of the distribution it was trained on and checking how it has generalized. That's a big nono in ML.

I still like the experiment though, it's fun.

The test tones used for training NAM and Tonex both contain sine sweeps and impulses which cover the entire audible spectrum, so I don't see how the sine sweeps used here could differ.

 

[yeky83]

At which point I was on the verge of giving up on modelling but then tried running my HX Stomp into the effects loop of an amp with no IRs. And the disorientation disappeared. I subsequently tried the Stomp with a PS-170 and a regular guitar cab and again experienced no adverse effects.

I mention all this because I wonder if it’s indeed aliasing that’s behind what I’m experiencing, why is that running the modeller through a regular cab corrects the problem. And also, is it the reason that people who may not suffer the extreme reaction I have but are subtly experiencing the same effect have a marked preference for using a modeller through a regular guitar cab.

Probably not related to aliasing since the Stomp is aliasing all the same into a cab (which to be fair could be masking the effects of aliasing I suppose...), sounds to me like your ear-brain just doesn't like playing guitar to the sound of mic'ed cabs.

 

[James Freeman]

No, the actual guitar signal.

I could try a brickwall filter at 19kHz on the processed DI amp sound, then train the NAM model.

 

[FractalAudio]

How do you not add extra compute if you want to bake in an anti-aliasing filter... wouldn't it have to oversample to do so?

You can't. The only way to avoid aliasing is to increase the sample rate. Double the sample rate and you double the computational cost. At a minimum. Often times this will quadruple the computational cost because it becomes O^2 operations.

Anything that does convolution-like processing (like a NN) will end up with O^2. I.e., if you double the sample rate the number of coefficients in an FIR doubles so you have to do twice as many operations for the FIR at twice the rate.

Anything that generates significant distortion should be oversampled (assuming 44-48kHz native sample rate) by at least 4x. That's an absolute minimum IMO. This would cause an increase of 16x computations.

 

[DLC86]

Probably not related to aliasing since the Stomp is aliasing all the same into a cab, sounds to me like your ear-brain just doesn't like playing guitar to the sound of mic'ed cabs.

Was going to say the same thing, I've experienced something similar and I think that's just ear fatigue induced by the emphasized hi frequencies in the IRs (it would probably be the same with a close-miced amp in headphones) plus maybe the lack of ambience/reverb in headphones

 

[imperator]

The test tones used for training NAM and Tonex both contain sine sweeps and impulses which cover the entire audible spectrum, so I don't see how the sine sweeps used here could differ.

All the way up that high. Hmm, if that's true then I have been proven wrong.

 

[DLC86]

I could try a brickwall filter at 19kHz on the processed DI amp sound, then train the NAM model.

Well, there should already be a filter: the anti-aliasing filter of your audio interface or plugin you're capturing

 

[fretworn]

Was going to say the same thing, I've experienced something similar and I think that's just ear fatigue induced by the emphasized hi frequencies in the IRs (it would probably be the same with a close-miced amp in headphones) plus maybe the lack of ambience/reverb in headphones

Agreed. The most obvious answer is that it is high frequency content.

Anybody that’s ever mixed for any period of time can tell you how the higher frequencies can wear you down when listening.

In fact, it’s a definite tell when you whip out an EQ to go deal with that nonsense.

It’s not hard to extrapolate then that aliasing creating shit up there is going to wear and work on your nerves.

 

[FractalAudio]

Well, there should already be a filter: the anti-aliasing filter of your audio interface or plugin you're capturing

Yup. If your interface is running at, say, 48 kHz, there will be almost nothing much beyond 20kHz.

 

[imperator]

I was assuming for guitar y'all didn't want anything beyond 10k? So there was an opportunity to filter a bit more

 

[DLC86]

All the way up that high. Hmm, if that's true then I have been proven wrong.

I just checked and both in NAM and Tonex test tones there's some unfiltered white noise reaching 24 kHz