Neural Capture: accuracy enhancements to keep up with competition (ToneX, NAM etc)

Quad Cortex Feature Requests
123

Totally agree - I love trying out stuff from the cloud, but modelled amp are more inportant. I wouldnt be surprised if Nural do go back and improve some models; theyve probbalybjust been focused on adding importand features such as hybrid mode etc and plug-in support (which i still dont ubderstand the fascination with, although i did buy the Rabea plug-in in the sale, but yet to try on my Mac - chise that one as want to mess with the synth function
*really need to experiment more with captured drive pedals on top of clean amp)

1 Like
124

As someone who reverse engineered the QC and the ToneX, I can say this:

  • ToneX is using an LSTM model, relatively light for inference but pretty demanding for training.

  • NAM is using WaveNET, still takes some resources to train within a reasonable amount of time.

  • QC is not using a neural network or any real ML really, not in the modern sense. QC is using a genetic algorithm using its DSP. This is very fast for inference and also very fast for training, that is why the capture process can happen directly on the QC unlike the two other options.

I think NeuralDSP could/should do this:

  • First step would be supporting NAM models in the signal chain, only for inference. So that we can start using them and they donā€™t need to care about training at the beginning.

  • Second step would be to implement the training part of these better models on the Desktop app, so that training doesnā€™t happen on the QC and we are all happy (easier for them to implement, they donā€™t have to worry about the limited QC hardware, etc).

  • Third step would be to implement their own higher def models (WaveNET, LSTM or whatever they like ) and training procedures.

Iā€™ve seen their software, itā€™s totally doable, especially the first step.
I think this would be the ideal solution.

15 Likes
125

The #1 enhancement to the ā€œcapturing processā€ I would like to see is noise cancellation (or a variant of it).

Iā€™m all in for the QC. Having had it for multiple years now, its an incredible tone shaping machine.

Ive gone through great lengths (time/experimentation) to get ā€œqualityā€ captures. Capturing ā€œoldā€ or ā€œvintageā€ amps can be tricky, especially if theyre noisy (humm, buzz, microphonic). I DONT fault the qc for the outcome. But i wonderā€¦. Before the qc starts its tone cycle, could it analyze the ā€œno sourceā€ signal first, before it captures? Creating this baseline could remove any unwanted sound in the new algorithm. This could be optional too (switch) provided in the capture menu.

The main difference in tone capture-wise (qc vs amp), is that with the amp the ā€œnoise is underneath the source (guitar) whereas with the QC-capture the noise is baked into the source signal (ie no play equals no noise/playing gets tone with noise). My 3 cents.

3 Likes
126

Itā€™s been a while since I posted this thread. Itā€™s refreshing to see folks converge on this one.

As some have suggested, even if the Neural Capture feature doesnā€™t progress much past what it is today (albeit I hope it does), maybe NDSP could open up & allow support for loading ā€œstandard architectureā€ .NAM files?

Anyway, Iā€™m back on the QC as of today as it fits my current needs better than anything else available at the moment. Putting aside my gripes with the unit (that lead me to sell the first one), if not the most powerful, itā€™s definitely the sexiest floor-modeler out there.

3 Likes
127

Neural if your watching, please add the ability to use NAM trained profiles on the quad cortex.

2 Likes