---

Do I include the boosters/speaker cabinet/cab sim in the chain?

---

Deliberately including speakers in the chain

You could experiment with including the speaker (sim) and even your booster pedal in the chain. If you can get the model to sound right, it can help you simplify your workflow. The developers noticed that it takes the heaviest training model and yet again it is no guarantee for a good sound.

With a bit of luck though, you might end up with one model that includes several components in one model; You’ll need less setup time figuring out the settings.

Live artists should keep their flexibility in mind when going for “cab included” models. A possible scenario could be a situation where you need to switch from “straight into the mixer” to playing over a real cabinet or vice versa.

Bad practice: two cabinets or no cabinets at all

Two stacked cabinets Mushy sound
No cabinets = Harsh sound!

When the speaker (sim) sound is a part of your chain on a pedalboard to go straight into a mixer, it will sound mushy going through a real cabinet. The double cabinet sound isn’t desirable, and I speak from rehearsal room experience when I say you probably won’t like it. 

Use a suitable input.wav file

Creating your own input.wav file

I took the liberty of adding my own section of long chords and fast chord chugging to the file. It makes the file longer but I noticed it offered a slightly better result on some of the high gain simulations. I have yet to prove its result but since some of my models will be used in a band that plays in “drop C# tuning”, it won’t hurt adding a section of that to my file.

Making the WAV files longer makes them larger in size, which quickly has an impact on the upload time when going through the training process.

---

Try to match db peaks and aim for a -6db peak

It could be changed by the time you are reading this but at the time of writing this article, the advice stands: Try to have a maximum peak of -6db in your target.wav file when you render. Having both input and target file on peaks of -6db *could* help the accuracy of the model. I’ve heard the team considered to handle this automatically during the modelling process at some later time.

This recommendation is made because the pregain can be adjusted up to +6db. This is not that crucial because the out gain can be adjusted in the json file produced by the training but it is of course a more clean practice to avoid that. 

Check the peak db when you create your render to know this peak. Most DAW’s show this value in the render window. Doing this, you mitigate some of the risk of your model being abnormally loud or silent.

Different training models; from “very light” to “heavy”

Don’t be mistaken, a “Light” training model doesn’t mean “low quality”

If you consider the lower CPU usage of this kind of model and its impact on your pedalboards for live use, you might be surprised with the result of the lighter settings. I’ve been using the Standard models so far and if you are using the model on a high end computer as part of your effects chain in a DAW, you can easily switch to the Heavy models if you want to. Then again, can you hear the difference between Standard and Heavy on your modelled amp?

---

What kind of models to deliver? Different qualities? Different gain settings?

The default setting, “Standard”, is the best options if you want to create just one model.

My idea is that the Community has more use of having different settings, all created on “Standard” quality than having just one setting in 4 qualities.

This is probably the most efficient idea if you want to deliver a different range of settings, like a low gain, medium and high gain setting. An amp specific mode or setting is always a useful one to include as well. (some amps have specific modes, channels and boosts). In that case, it might become a tedious job if you want to include different “quality” levels for each setting.

When you are modelling just one setting on the amp, the best Community service would be to deliver the whole set of qualities, from Lightest to Heavy.
When going for two models; I’d go for Standard and Lightest or Heavy and Light (both having a setting between them.)

---

What does “validation patience limit reached” mean?

When training a model, the process can stop before it has reached the chosen amount of iterations. You’ll get the message “validation patience limit reached”. This message means that the model is not getting any better and instead of running until the end it stops; it isn’t learning anything new anymore.Basically saving you the time with the same result of doing all the iterations you selected.

---

Using your MOD Audio device for re-amping a pre-amp/pedal

!! MAJOR WARNING !!
* NEVER use the speaker output of an amp to go into your MOD device. It is literally powered up to physically drive speakers.
* If you want to model an amp, you might prefer to limit yourself to the PRE-AMP SECTION.
The output of the pre-amp section can be tapped from the “FX send” (or similar) output. This has low power and it’s suitable for sensitive electronic gear (since you use that one as effects loop of an amp).
* If you insist on modellign the POWER AMP section of your amp as well, you mus use specific gear like a loadboax, physical speaker sim with low powered output etc.
* This guide is aimed at people modelling only the PRE-AMP section or pedals.

The process is simple

> I press play on the file player
> I check the levels on the little displays of the gain plugins right after input 1 and 2. 
> I adjust the outgoing level on the device I’m modelling to match the level with the dry signal somewhat, making sure it isn’t clipping and showing somewhat the same behaviour has the input, level wise

> I make the recording by stopping the file player, pressing record and starting the audio player again. 

> Stop recording when the file is done. Don’t worry if the file already started over again, since your recording includes both dry and wet track, you will always have the same file length.

Creating a pedalboard for re-amping

Plugins to use in this process

After you upload the input file to your MOD Device, you can use it in the File Player.

I use one of Brummer’s file recorders to record the target.wav AND an unaltered input.wav. This plugin can be found in the pedal library for you to download via your Mod device’s interface. You can use either the 4-input recorder or the regular stereo recorder and let the dry (original) signal some in on L and the wet (amped) on R and split them later in your DAW. I find Audacity the easiest tool to split a stereo track to two mono tracks afterwards.

The TinyGain plugins are easy to use and feature a little display that helps in having an indication of the levels

---

ESR Values: What is the accuracy of your model?

ESR= Error to Signal Ratio.

The ESR is a value that expresses the accuracy of your model. It is a number ranging between 0 and 1. The closer the number gets to 0, the higher the accuracy of the model. 

ESR quality guide

0	Absolute zero, impossible to achieve
0 -> 0.05	Great success
0.05 -> 0.15	Good, accuracy still within the comfortable limits
0.15 -> 0.2	Slightly noticeable difference in some cases but probably still a very usable model
0.2 -> 0.35	Make sure you do listen well. This will be a usable model but it is possible the amp sounds more like a “sister amp” than the actual amp.
0.35 ->0.5	Could be better, it might pay off trying again after checking all the steps again
>0.5	Make sure all your parameters were taken care of, as these values. Chances of having a usable model are slim
>0.9	Failed, check the levels and alignment or try re-amping again from the start.

High ESR? Check for misalignment issues

The current version of the training algoritm automatically detects the two short peaks in the beginning of the input.wav file so misalignment should not be an occuring issue. If you keep getting very high ESR values (0.7 and higher), try the procedure below to manually align input.wav and target.wav files

Several Community users noticed that misalignment issues can occur when creating the target.wav. This can have several causes and usually latency is the cause of the issue. Latency can occur if you have digital-to-analog and analog-to-digital conversion, latency on effects chained in your DAW etc.

You can check this by importing both your input.wav and target.wav in your DAW. Zoom in as far as you can and you will end up looking at your signal on sample level. It will look like this:

(screenshot by Community user Itskais, member of the MOD Audio team)

The screenshot already zoomed in on the absolute beginning of the input.wav file. Above, you see the input.wav and below the target.wav. 

Do you see the 2 short peaks? Those 2 pops are especially created to help the algorithm align the signals. It works like a clapper used in movie production, to align image and sound tracks.

User Spunktsch, part of the MOD crew, sometimes adds a standard “click” sound in Reaper. Look how the clean sine wave, consisting of samples of various heights, is being changed into a square shaped sound in the target.wav. 

(screenshot by Spunktsch)

If you look carefully, you’ll notice the green line denoting the start of the cycle. If you notice that your files. 

User Jandalf on the Forum came with this handy idea to align waves: he could fix it by aligning a zero crossing on both signals. Meaning: he looked for a spot  where the input signal switched from negative to positive signal values. Then he took the target signal and shifted it forward until the zero crossing of it was at the same sample as the one on the input signal.

An example of bad alignment

Zoom out a bit again. This screenshot shows a clear misalignment. In the target.wav, the first “pop” still needs to happen when the first one occurred. Try adjust the position of your target track to the waves align.

Identifying the wave shape for precise alignment

Mind you, identifying the start and end of a wave isn’t all that easy. Take a look at the screenshot by Itskais again. This is the result of a successful re-amp process and considered a valid match. Although the target wave seems to start earlier (going down a bit), most of the stronger bit above the line matches up…but not completely. 

Introduction of harmonics in the distortion, behaviour of he amps, etc and other aspects of the process can make it harder to identify alignment but notice the sheer difference between the screenshot below, where you’re talking about samples while in the example above, you are talking in terms of time

---

Epilogue

“Imagine having an algorithm learning how an amplifier or pedal works based on comparing a dry and a processed sound… You could sound like every amp”

LievenDV, August 1999

Fast forward almost a quarter of a century to a time where the terms “AI”, “Neural Networks” and “Machine Learning” become trivial though we don’t always understand how it works.  Is that an issue? Not at all, since the bright people of AIDA and MOD Audio have made the technology accessible for mere mortals and guitar gods amongst us. 

I’ve heard (of)modelling before and some brands made their name by selling pretty accurate models and an expensive way to model gear yourself. The concept isn’t new but it has been a while since a disruptive player emerged on that market. 

When NAM caught traction, you no longer needed expensive hardware but you still rely on the processor power and DAW software on your computer. Modelling became accessible and very affordable.

The next breakthrough came from AIDA dsp and MOD Audio, offering open source technology on relatively cheap and mostly a very portable and gig worthy device, the size of a big guitar pedal… and here we are, living the dream I had in 1999!

---