An MIR analysis of real amp versus captures and profiles

So I’m no expert. Any of this is free to be questioned or replicated and puzzled over.

First some background on the mel-spectrogram:

Understanding the Mel Spectrogram

A quote from the article that gives us a simple understand of the mel spectrogram:

Studies have shown that humans do not perceive frequencies on a linear scale. We are better at detecting differences in lower frequencies than higher frequencies. For example, we can easily tell the difference between 500 and 1000 Hz, but we will hardly be able to tell a difference between 10,000 and 10,500 Hz, even though the distance between the two pairs are the same.
…
…
A mel spectrogram is a spectrogram where the frequencies are converted to the mel scale.

IE: A mel-spectrogram is the spectral representation of an audio signal scaled against the mel-scale, which more closely represents how human hearing perceives frequencies.

So I’ve never been entirely convinced by the accuracy of the Kemper Profiler. I could always hear differences, and after purchasing six of them since 2011, and never being entirely happy, I decided I would wait to see what the company does in the future. Similar story with the NeuralDSP Quad Cortex; I purchased it and actually really enjoyed the capture functionality. But there was so much about the device that I didn’t enjoy, that I quickly sold it and now I’m looking at them from time to time to see where the product goes.

But I still had some audio recordings left over; recordings of a real amplifier and then the Kemper and Quad Cortex equivalent profiles/captures.

Please note, I am only talking accuracy here. I’m not making good versus bad tone claims. I don’t own a Quad Cortex nor a Kemper anymore, so I have a very small data set, so none of this should be taken as gospel or an ultimate conclusion of anything. I just find it interesting.

So Here’s what I did:

First I normalised the audio files in the DAW, to minimize digital level differences (not the same thing as balancing signals using your ears!)

Then I took a mel-spectrogram of each file. I then converted the data from power-spectrogram to dB.

For each modeller signal, I added it to an inverted version of the real amp signal; essentially a polarity inversion that you would do with audio in your DAW, but against the mel-scaled spectrogram>dB representation.

And these charts are what you get.

Mel-spectrograms of the source amplifier recording subtracted from the capture signal in question

Effectively, for each frequency band…. the closer to zero you get, the more accurate the modeller signal is to the original signal. Where the colour is pink or red, those are the biggest differences compared to the original signal.

So from a high-level macro view, I observe some things:

• The Kemper has consistent differences in the low frequency range.
• During the palm-mutes in my riff, the Kemper suddenly snaps back into “very accurate” from 15,000hz down to about 7500hz.
• But on the open notes in the riff, this doesn’t happen. There are quite a lot of differences on those notes.
• Also on the palm-mutes you do see interspersed red chunks throughout the frequency range. This also happens with the QC.

If we zoom in on the lower frequencies:

0hz to 7000hz of the same mel-spectrogram graphs

It may look like a load of noise. But I’m seeing some patterns in there that are quite interesting.

There are far fewer dark red chunks on the QC lane (3rd lane if it wasn’t clear) than the other two. Below 1000hz. So let’s zoom in there….

So look at frequencies 0 through to 500hz. The colour-banding here should tell you about the accuracy of the signals versus the real amp.

The Kemper has a much higher base level of pink and red in the 0hz to 500hz, across the file. This shows that in that frequency range, the Kemper is less accurate than the QC.

IK Multimedia ToneX

So now for IK Multimedia ToneX. Unfortunately I made the Kemper VS QC clips in 2021, and don’t have the DI anymore. So this is a completely different riff. But it has loads of palm mutes, strummed chords, and ghost notes and such. We’re still comparing a real amp to a capture of it.

ToneX version of the same process

Ignore the amp line at the top — actually no, don’t. What that is showing is that the inverted amp and amp data completely cancel each other out (constant stream of zero’s in every single band) which means my analysis is doing the right thing.

There are differences of course. But fewer instances of bright red and super dark blue.

Here’s the frequency spectrum and the mel-spectrogram analysis for a ToneX capture of my Diezel D-Moll:

A frozen spectrum captured in Fabfilter Pro Q-3 EQ plugin

The Diezel D-Moll capture I made with ToneX, using the same mel-spectrogram inversion process

You can barely see the data, it is so close to the noise-floor. This means it almost phase inverts against the original amp recording — ergo; they match very closely.

Finally, wrapping this up with a comparison of advanced and fast training methods:

Advanced mode versus FAST mode

Fast mode versus advanced mode.

Advanced is the 2nd graph, fast is the 3rd graph. The 1st graph is just my purple ‘zero’d out’ real amp graph. It shows my inversion formula works correctly.

Definitely noticeable high-frequency differences when using fast mode. But still reasonably accurate.

My advanced training times are about 15minutes since the last update. FAST mode was 3minutes.

BTW, I’d like to point something out:

We are better at detecting differences in lower frequencies than higher frequencies.

Since this is true, it means it is even more important than I ever considered, to make sure you get the low frequencies correct when designing these algorithms.

There is a bit of a fallacy lurking around that the guitar is a mid-range instrument. It isn’t. It’s a broadband shotgun of noise, shaped by filters. That needs to be the mentality when going into amp modelling and capturing, IMHO.

Again, this isn’t exhaustive and shouldn’t be taken as any kind of conclusive truth. But it is interesting to see!

Conclusion?

My ultimate takeaway from all this…. There is a threshold where amp capturing becomes acceptable, and this is different for all listeners. Many people are fine with the Kemper and don’t even hear the differences, and they swear blind that others are just being overly precious or difficult. And those others cannot believe people cannot hear or do not care about the differences with a device that purports to capture your real amp and digitise it.

So there is a threshold of acceptance, and I think the aim of the game isn’t to get a 1:1 digitised copy, because it is probably impossible. Instead the goal should be getting as many people over that threshold line as much as possible — and THEN focusing on user interface, user experience, and feature set.