I’ve been trying to capture my Marshall Silver Jubilee 2525H Studio using my Mod Dwarf, the NeuralRecord plugin and the AIDA X platform. The problem is that the output is far from good, and I just can’t figure out what I’ve done wrong.
The amp was connected to a cab, I hooked up the DI output of the amp to a Imp2 DI Box, and then connected the low output of the DI Box to Mod Dwarf’s input using a XLR-TRS cable. The output of the pedal was connected to the input of the amp. I then used the Neural Record plugin to run the input.wav and record the output of the amp.
I normalized the amp output to 0db before running the AIDA X training, and the initial ticks of the target were aligned to the input. While the two audio waves look aligned through most of the input lenght, it’s possible to see some spots where they are misaligned, but I can’t align them without ruining other spots, so that’s probably where the problem lies, but I don’t know what I can do differently.
While the ESR values I got where around 0.05, the simulations sound too direct-to-board-like, they’re too thin and trebly and don’t resemble the target when I hear them on the results section. I’ve trained using standard size and 400 epochs and it always ends around 200 epochs.
I’ll leave both the input.wav and target.wav as they were extracted from my Mod Dwarf so anyone can’t take a look and teach me what is wrong or train their models using those files to see if they can get better results than me. I really appreciate your help and the dedication of the community, being a Mod Dwarf owner has been a very fun ride!
I’m at work so can’t really check right now
The DI output of the amp, is that with a speaker sim or without?
if without speaker sim, have you tried several cab IR’s when testing the model?
It’s without speaker sim. I know the result will naturally sound like a preamp direct to board without using any speaker sim, but I’m comparing it to the target that also doesn’t have any speaker sim and they’re always soundng considerably different, and the difference signal has a lot of higher frequencies that were asbent on the target too
While creating some models myself, I had a few that I just could not get to work properly.
What worked (sometimetimes and also partial results) is trying a different input file. I took the original one and added some “genre specific” things like some down tuned fast and slow chugging etc.
I even captured some VST’s BUT with some EQ tweaks where I toned down some of the higher spectrum to eliminated some of the “woosh”.
Yet, there are models where I ended up in situations like you
Sorry I can’t be of any more help on this
Thanks a lot for helping!
I tried capturing a UAD Lion from a friend, I think sharing it might be against the rules but the capture turned out really great, very similar and usable tones, so I think the jubilee might be one of those bad situations. I’ll try recording more guitar parts for the input file in the future!
Thanks a lot for helping! If you get any free time and really feel like doing it, please take a look at the target.wav I shared to see if I really didn’t do anything wrong! Hopefuly there’s something wrong with what I’m doing lol
After some further testing, I was able to get this 0.017 ESR model of the amp on Standard size, the target vs model snippet comparison still has noticeable difference but this is the best I was abe to get, I’ll test it on the dwarf tomorrow and upload it to AIDA X Cloud if it does sound great through cab sim
On another note, what’s the difference between this .json and the .aidax needed to upload the model to aida x cloud? Can I just safely change the extension to .aidax, is it merely a convention issue?
Marshall 2525H LSTM-16-0.json (53.4 KB)
as @LievenDV noted the standard input.wav sometimes isn’t the best to capture some amp types.
You get better results with a few guitar riffs added or use another one.
it’s just a renaming of the file extension. Provides an easier way to filter and share files.
thanks, I’ll check it out sometime and run it through our process.