Model only makes signal ...more silent

Hi guys,

I’ve been experimenting with the workbook and I’ve been able to render some results but now I keep getting a result that only makes the dry signal more quiet. It was supposed to be a high gain plugin setting in reaper.

sample count, hz, format…they all fit.

I tried it 3 times with different folders and different renders from reaper.

device = MOD-DWARF
file_name = roots-v2
unit_type = LSTM
size = 16
skip_connection = 1
 20% 61/300 [02:22<09:14,  2.32s/it]validation patience limit reached at epoch 62
 20% 61/300 [02:26<09:32,  2.39s/it]
done training
testing the final model
testing the best model
finished training: roots-v2_LSTM-16
Training done!
ESR after training:  0.9874507784843445

In this step, the “predicted signal” is virtually the same as the dry.

When I load up a (new) Reaper project and use the Aida VST, the plugin without cab IR but just the JSON freshly created only makes my signal more quiet.

Symptoms are basically the same as mine (Validation patience limit reached at epoch - #6 by ignis32) You are not alone in this boat)

P.S. Are you using google colab environment? (I am asking because I was blaming my local environment a bit, but if you get the same in real colab cloud, I can focus on the input files instead)

Was thinking in that direction as well


After that attempt, I did another try with another set of sounds based off a different plugin re-amp and it rendered my best result of the day (NRR-1 high gain channel)

it’s not that it’s “completely broken” or something :smiley:

Hey guys!

An ESR value that’s close to 1 is pretty much awful, which confirms that the final model is not doing what it’s meant to do.

If the problem is sill persistant, maybe I can have a look at the input files and potentially help see what the problem is?

@LievenDV Maybe this response can be insightful for you too!

While everything is there, taking a closer look reveals some spacing

This is one of the pops.

Would the latency be the cause and could it be the cause of me re-amping a VST amp in reaper?

I was thinking of reamping with my Dwarf today:
Play an audio file through my output and capture the signal that went through a pedal with my input.
Knowing there IS latency when using the Dwarf, this is probably to such a great idea?

I can’t exactly determine how much latency is there between the 2 files in the image. Also I can’t really see how a simple amp can react in the way that it’s shown in the image, the target reactions seems to be too long and complex for just a simple pop in the input (maybe I’m wrong here idk).

Regarding reamping with the dwarf, I believe this can def be possible as long as you manually align the input and target files afterward ^^
Hopefully this step of “Manual alignment” can be implemented to be done automatically in the training notebook in the near future.

latency on sampel scale before it even starts to “move”

Latency to the highest peak

Latency to end of the response

It’s a very high gain affair but there is no other plugin or effect in play here

yeah worth the try and then checking out whether I can align it properly.
We must explore all workflows that seem intuitive

1 Like

Maybe you can try to manually reduce the ltency between input and target files, and see if it makes a difference in the training, seems to me that there is indeed more latency than other reamping files that I did training with.
Also, please make sure that input and target files have the same length in samples. (this step is normally already done automatically in the colab notebook, as long as the difference is no more than 3 secs)

Let’s see how the pedal thing works out ^^ which pedal is it?

Yeah, this is certainly an important thing to check, especially if people are modelling VST plugins.
I’m going to test the same clips again but with a matching in timing I think is more accurate and post the results here.

I only upload sample perfect clips, 0 difference total sample size.

I have several lying around I already posted online for selling; could just as much try to model them before they go :smiley:

Strymon Riverside
Wampler Sovereign
Wampler Velvet Fuzz


I was able to bring the ERS back to a lower value but still pretty high.
It started to sound like a high gain amp… but not like the original.
It didn’t give me the silent treatment anymore though.

confirms that aligning surely helps though.

75% 225/300 [08:12<02:29, 1.99s/it]validation patience limit reached at epoch 226
75% 225/300 [08:15<02:45, 2.20s/it]
ESR after training: 0.6150435209274292

I’m blaming the VST amp.
a different VST amp does a lot better


I just achieved the exact opposite of desired behavior.

While my first go at the Strymon riverside gave me a ESR:0.004 or something, I just nailed 0.9969 :confused:

ESR after training: 0.9969747066497803

This was actually AFTER re-aligning the files to compensate the latency.
(before that it was 0.96 instead of 0.99)

Is this due to a near-perfect phasing where I canceled it out or something?
It renders…silence :stuck_out_tongue: