NAM: Neural Amp Modeler

SwirlyMaple said:
Is it just me, or does it sound like there is dry signal bleeding through all of those clips except the nulls? :idk

I will say the nulls are impressive.
Click to expand...

It's not you - that's exactly the point :) It's not a dry signal, it's a capture of a clean and high gain amp combined in parallel (the combined result is what's captured and compared to the actual combination). It's not supposed to sound good, it's just supposed to illustrate NAM capturing blended signals.
 
northern_fox said:
It's not you - that's exactly the point :) It's not a dry signal, it's a capture of a clean and high gain amp combined in parallel (the combined result is what's captured and compared to the actual combination). It's not supposed to sound good, it's just supposed to illustrate NAM capturing blended signals.
Click to expand...
Oh! Ugh. I'm sorry. I'm an idiot. I even knew that's exactly what you were doing but I guess I didn't expect clean to mean that clean. :rofl
 
James Freeman said:
Blending sounds legit here.
Goes to show the major benefit of neural network captures over the previous Kemper type preamp-matching.
Click to expand...

I was honestly expecting it to be a lot farther off, especially in the first clip because I cranked the bass on the clean amp too. I haven't tried blending gain + gain yet, mainly because it's harder to really "hear" whether it's correct afterwards - but maybe I'll try that next. I also did not time align the separate tracks with each other prior to rendering, though they were really close anyway.
 
Mike Oliphant said:
Another thing to do would be to line up your input and capture files, and see how much longer it takes your capture to fall to the noise floor.
Click to expand...

I happened to have the capture output for a high gain amp (Diezel VH4 - captured by Petr Canov). Here it is alongside the input wav:

AmpTemporalResponse.png


The files aren't properly aligned, but you can see one of the alignment "blips" in the v_1_1 training file at the top. It is just a few samples long. Then look at the size of the response - over 2000 samples. It is easiest to see here, but the the temporal response is similar for the rest of the file.
 
I've tried with Resistive (flat) and Reactive loads, and also just the Preamp from the FX Send, I see this temporal response with every output so it's not the output section of the tube amp.
Maybe it's the big Choke in the power supply of the amp, but it's well filtered and I'm recording at quiet bedroom volumes, so I don't know.

The present response depends on input signal from the past combined with the present input signal so it's a time variant response, it cannot be captured without some degree of error, such is the nature of tube amps.

impulse.png
 
James Freeman said:
The present response depends on input signal from the past combined with the present input signal so it's a time variant response, it cannot be captured without some degree of error, such is the nature of tube amps.
Click to expand...

To the extent the time-based response is *consistent*, and does not have dependencies beyond the temporal capabilities of the model architecture, it should be able to be captured.

Out of curiosity, what does the temporal response of the NAM model look like? (you can see it by running the training file through the NAM plugin and recording the output)
 
Mike Oliphant said:
Out of curiosity, what does the temporal response of the NAM model look like? (you can see it by running the training file through the NAM plugin and recording the output)
Click to expand...

This is what I'm seeing with the NAM model corresponding to the Diesel capture I posted above:

NAMResponse.png


Looks like it is doing a pretty good job of matching the initial temporal characteristics (although, interestingly, it is noisier than the original signal). It can't match that second transient (not sure what that is, to be honest), as it is likely outside of the receptive field.
 
Mike Oliphant said:
This is what I'm seeing with the NAM model corresponding to the Diesel capture I posted above:

View attachment 6483

Looks like it is doing a pretty good job of matching the initial temporal characteristics (although, interestingly, it is noisier than the original signal). It can't match that second transient (not sure what that is, to be honest), as it is likely outside of the receptive field.
Click to expand...
Yea that looks odd, you would think if that second Impulse was presented to nam as an input it would recreate it as well.

I guess a question might be if you modeled the amp as an autoregressive process, how many delays you need before the coefficients become negligible. There's obviously some balance between accuracy and latency. You could use a transformer but they might be too beefy for real time use.
 
I've asked question on Discord and they were very helpful, particularly the admin there.
I think people are still very hyped about this and are sharing captures of anything and everything they can put a cable through.
 
Orvillain said:
I had to leave the NAM group on Facebook. Too many swivel eyed C*nts on there.

I'll share whatever captures I have here soon.
Click to expand...

Mongillo19 said:
They do love NAM to a fault it seems. Really isn't much room for conversation unfortunately just from what I've seen
Click to expand...

Funny [not] - in "the other place" - after praising NAM for what it currently is, I also mentioned / suggested some points of criticism and where I thought it needed to focus to get even better - received multiple responses from quite a few people which were bordering on deliberately insulting, to outright nasty - others who expressed similar points, were equally "attacked".

Still there lurking [FB] under a different name, but its been pointless for a while now to post anything there with any semblance of constructive criticism / suggestion / recommendation / comparison.

Ben
 
Last edited:
You must log in or register to reply here.
Back
Top