Campaign: Share also you input and target.wav too

Neural Modelling Model Training
#1

Hello, everybody.

As we can noticed, the amp modeling with machine learning stuffs is a hot topic. We have NAM, Tone X, AIDA X and others. The community is doing a awesome work getting their great sounds accessible for everyone.

I noticed that some of our friends are requesting converters from or to NAM or Tone X. As you maybe already discovered, this is not technically possible, because they use different “internal structure” (different machine learning models), also Tone X is closed source.

Other are requesting a NAM implementation for MOD Dwarf. At the current state, isn’t currently possible: although AIDA is implemented a generic way, that enables load future improved models (LSTM that you can see on some .json files as an type of model), NAM files doesn’t use this generic pattern. Also, NAM models are generally CPU heavy for MOD Dwarf.

I would like to you think: is the AIDA DSP the current state of art? Maybe/probably. But… Until when it will be? Do you think that someone will improve it or maybe create an alternative that is incompatible with AIDA? Maybe the next evolution would improve the sound similarity, maybe it can consume less CPU, we don’t know. But one thing we know. The evolution will occurs.

So I’m asking you, my profiler friends, would you like to try the future pedal? Of course, right? But do you like to use the previous captures, probably with a better quality with the future technology?

If you answer yes, so please share the input.wav and target.wav files too. So, in the future, new AI models could use them for making amps and pedals compatible.

An analogy: Sharing the input.wav and target.wav is just like share the source code of a video game. Maybe you did the software for running on Nintendo Switch. But someone can use it for generate a new “executable” for XBOX One or PS5, or can also improve and generate a better polished for Nintendo Switch too, like with anti-aliases stuffs.

Now there are a lot of NAM model files, if the modelers shared the training files (input.wav and target.wav), now we could use them now for training and use on AIDA too.

So, for the technology future improvement, please share also the input.wav and target.wav.

Edit: I rewrite it for better clarification and add an analogy.

10 Likes
#2

Sorry, I don’t understand:

  1. why do you need the input.wav? That file is always the same. Or not?
  2. On forum you can’t upload .wav
#3

Hello, @Teuvosick.

Non necessary. Users can use different input.wav at some cases. An example from here:

It also facilitate when someone decide to retrain a new model, getting the input and output at the same place. For instance, if someone of NAM community would like to make a model from a AIDA dsp, they maybe wrongly think that is necessary to use a NAM input.wav (I’m supposing that they use a different input).

Ok, I don’t have an answer for this. It will requires most effort of community on share the files using external sites :confused:

#4

exactly. We have been talking to some People over at the NAM discord. Someone provided a few input and output pairs. One issue I see posting those .wav files is that some parts may be not your own and fall into copyright claims.

But other than that it’s a matter of time and user adaptation. ToneHunt.org is a good example for that.

1 Like
#5

This thread kinda leans on my recent thread (with no inpui) about best practices and conventions Modelling: best practices & conventions

Sharing input (if diffeent than MOD’s) and output sounds interesting for ohers to give it a try but I think the most valuable would be trying to discover which files work best for which case? If it proven that the provided input.wav is the best for certain genre of amps?

1 Like
#6

When @Teuvosick questioned how to upload, I thought about Tonehunt. But I don’t know if is feasible ($$) because it requires more cloud storage for something that will be probably hardly ever used.

Hey, @spunktsch. I read on some post that you and other are talking with Tonehunt team, right? If they open space for AIDA X files too, would be awesome if they add too the “AIDA X wasm player”, so the users would listen just in time the result :star_struck:. And it will increase the focus on AIDA X.

@LievenDV, maybe I wasn’t clear. I wasn’t talking for users share differente input.wav for other users use them for its own equipments. I’m trying to explain that if you share your input.wav and target.wav (training files), when someone create a AIDA X2 or NAM 2 or other technology, we can use the training files for generate a model for this new technology.

Now about the discussion of using custom input.wav:

On machine learning, the training data is some of the most important things.

Maybe there exists an audio that will maximize the training process of a specific effect (eg, amp or guitar pedal, or similar) for a specific model. I.e, would have an audio with if you can use it for create a model of a specific effect that create the lowest ESR with the lowest time for training. But the audio for this effect will not have the same behavior for another effect, or another model.

I think that doesn’t exists on literacy papers that explore with data could perform the best results on this guitar stuff area. But, my bet is that the current input.wav is good/excellent for most of the cases, but maybe it could be improved at the someway that I’m not discussing on my original post :slight_smile:

2 Likes
#7

ah yes,
We’ll need some kind of repository then.

check!

1 Like