What do you think of this idea? Is there a chance that the MOD Duo X could be a guitar amp profiling device, similar to a Kemper or the Quad Cortex? Is it possible to implement this? Hardware wise, I think the MOD Duo X should be able to do this (may be except for some additional mic pre’s)? It probably has the required computing power? Wouldn’t it be great to capture the sound of our guitar amps with the MOD Duo X?
best regards
Warning, this is not a direct answer to your question.
Take a look at this thread for some (at least partially) related thoughts and information Non linear modeling of real amps using the MOD
Theoretically it may be possible, but needs the implementation - which can take a bit some time, depending on how this feature would be relevant for the community.
Anyway, there’s a big step into making this more possible coming with the version 1.10! That is saving files inside the device…in other words, you would be able to save the impulse responses of the amps.
But this could be a first shy step, since all the infrastructure of a well developed plugin to do this would still be required to be build.
An intermediate step could be to do the profiling on a more powerful machine like a PC, but to be able to use the profiles on the Mod Devices.
But look at the thread mentioned above about this.
Read it more carefully.
It seems that the plugin might be there…so together with the files handling system (and of course possibly some development work)… voilá 
It may be faster to get it done.
I will save this on the suggestions just to not lose track 
Ability to read IR profiles (even better create own profiles) is a must have feature for a guitar-oriented device. Presets sharing section should be updated too in order to manage needed copyright-free IRs since commercial IRs couldn’t be freely shared. I think this would add some complexity on the legal side.
On the technical side, just keep posted because soon you may have nice news about IR profiles
No need to keep suspense.
We are trying out a few things in regards to IR.
The quickest way we had to get something working was basically to grab the code from the cab simulators and make it generic so it loads user files.
While this works fine for small wav files, it is not suitable for big or even medium-size reverbs.
Still, it is a start. Plugin is already in beta for those that are in v1.10 already (which means anyone with a Dwarf, plus the few doing internal testing)
We investigated which IR plugin would be best to the platform. Most that exist are not well optimized for ARM.
Best results from our testing is using LSP Linux Studio Plugins Project
The developer has heavy ARM optimizations, adding even more each new release.
This plugin has a lot of controls that are not exactly needed for MOD, so we will hide them away.
While we have this plugin running on the platform now, no other work on top has been made yet. So you won’t see it (properly) in the store just yet.
Hi @falkTX, this is great, however IR loaders are one thing and profilers are a different. A profiler like the Kemper uses a different technique to capture the amp tone so that it can “reproduce” also the harmonic distortion. Is a patent-pending solution unfortunately. Recording a simple impulse response of a non linear device is non-sense since the convolvers only work with “space” or phase delays -> result is sort of equalization with some artifacts. I see some IRs taken from devices such as channel strips or tape machines, well from what I know of the theory behind convolution they’re non-sense stuff. Warning: I am not saying they couldn’t sound “good” I am telling that they’re not achieving their initial intention, to capture the tone of the device. May propose a different direction? With a different technique instead? Profilation is somewhat an outdated story already since “the guys” now want neural techniques applied to amp simulations (hell ya!). It was mostly an obscure thing for me since I discover this. What do you think? Is there anyone interested in this?
On the topic of neural networks, I did try a guitar-oriented plugin some time ago. Supposedly it creates realistic simulations of Amplifiers and things like that…
While the results were very nice, using it had a very big impact on CPU. On low buffer-sizes, even my desktop PC was being pushed a fair bit.
Related discussion Investigate reducing CPU usage of current WaveNet implementation. · Issue #13 · GuitarML/SmartGuitarAmp · GitHub
The link I’ve proposed runs on a Teensy. In the repo there is the c code that runs on the Teensy. There is also a youtube demo with the neural network starting from 1:26". I think if it runs on the Teensy should be ok enough for the Mod, isn’t it?
WaveNet is a super intense model. There are neural network approaches to pedal and amp sim that use lighter models based on LSTMs. From the papers I’ve read, there’s not a huge difference in sound quality with LSTM modeling. This is a topic I want to investigate further one day, and I even have a dataset for it. Hoping to come up with a solution that works on the MOD platform, even if the profiling part has to be done somewhere else.
I would also be interested in trying out the chow pedals. Not all of them have LV2 builds yet (or full cross platform LV2).
Yes, Chow tape is very nice: GitHub - jatinchowdhury18/AnalogTapeModel: Physical modelling signal processing for analog tape recording
He has also an interesting phaser: GitHub - jatinchowdhury18/ChowPhaser: Phaser effect based loosely on the Schulte Compact Phasing 'A'
Let’s make it simple. Can we name which type of neural network is used in Neural DSP products? Sound quality is heavily affected by the data set. In the papers I’ve seen, they all rely on a custom and very generic data set which (in my opinion) is not adequate. What I mean? Poorly played and not focused. Focus is important since when you’re using neural networks for amp modeling you want to focus on a particular playing aspect for example rhythm. So I would train the network to be very close to the original amp when I play, let’s say, power chords with a heavy attitude (on a 7 string guitar for ex.). Another training may focus on funk rhythms. In this way user would have an amp for the exact part he needs to play. I don’t know which network is used by Neural DSP, but by inspecting their patch profiles they’re combining three amp channels (clean, crunch, rhythm and lead). It’s very likely they trained their networks both on channel type (with various knob setting) and playing styles.