With several players in the amp/drive pedal neural capture game now, I’m hoping we also see equivalent capability for modulation/time based devices. Got some analogue pedals with funky filtering that I just can’t go without currently but would love to have them in my Duo X natively.
Given the sheer demand for the current tech, I feel, however, the next step is surely creating full models of amps/drive pedals from multiple captures and having parameters function as per the original device.
Anyone else feel the same, actively working on modulation captures or have a different outlook for the ‘next step’ in the capturing roadmap?
From what I understand this is already possible with “conditioned models”, where (as far as I understand) part of the training is made dynamic by exposing 1 parameter to the training side and doing iterations over a few steps of that parameter.
The RTNeural engine as used for the upcoming aida-x plugin supports this, and I have seen a few projects out there making use of it. It is typically just 1 or 2 parameters, not more. Maybe because it takes too long to train? Don’t know myself to be honest…
Anyhow I think there is potential to have something like this on the MOD platform eventually.
Do you have something in specific in mind? Putting some ideas out there could be the trigger to push someone to want to try this.
We are preparing some documents/tutorials to make it easy for the community to create new models, though initially without any parameters. Would be certainly be very interesting to try though.
the next step
Anyhow I think there is potential
we have the theoretical limit of 65535 parameters for conditioned models. This is the number of channels in a wav audio file. However, this would also impact training time. If I want to parameterize the gain, I would need at least five different audio recordings, one for each value of the knob, like 0.0, 0.25, 0.50, 075, 1.0. If I add the volume parameter, I don’t have another 5 tracks, so 10, but I would have 5^2 = 25 tracks. So you see that is somehow exploding pretty quickly. Also with current network models CPU would be too heavy only for a pedal. But the thing I like about this stuff is that it scales down. So we could experiment in future with lighter networks that are suitable to model less complex stuff like “Funky Filtering”. I also like those kind of filters on Bass & Guitar.