Capture your gear for free with TONEZONE3000

[[Nathan]]

I had a little bit of time last night to reamp and applied some of the recent training info from this thread. I thought I'd share for anyone who finds it interesting.

I put a bunch of NAM files in the following folder:
https://drive.google.com/drive/folders/12ikGAk-5YSD1NhHzzDIdBO0mWHTIeark?usp=drive_link

I made captures of 4 amp channels:
-Mesa Dual Recto Multi-Watt Red Channel Vintage Mode
-Mesa Dual Recto Multi-Watt Red Channel Modern Mode
-Peavey 6505 1992 Original Lead Channel
-Vox AC15C1 Normal Channel

With the Wet/Dry method, I trained the files in the following ways:
-Standard
-xSTD code with default learning rate/decay
-My training file with some higher freq sweeps a la the 50K file + xSTD code as above

At minimum, there's 12 captures you can cycle through to hear how the different training is affected with my setup.

Thanks for the captures, made it easy to quickly test it all out, gave it a quick 10-15 min spin on the Red Mesa

MB DRMW Red Mdn Test STD
I took this as the standard method. Seemed to have a touch more low mids

MB DRMW Red Mdn Test xSTD
Same as the STD but with the tiniest shift from less low mids to a treble kiss, makes palm mutes a bit tighter

MB DRMW Red Mdn Test NewInput xSTD
Another notch in the treble direction over the one above. Somehow feels a touch louder or gainier than the others but I'm talking 0.5%, probably just an EQ thing

From a players perspective just sounds/feels like different options or someone nudged the EQ by 0.2 in the mids or treble. Would be interesting to see how the new stuff NULLS against the original amp. Either way it's been a good quick test just to see how it plays first hand. My initial assessment is same same but mildly different, not world changing at all. If there's any benefit to all of this it's probably on the mixing side of things.

As a side note the embedded input spec is great. At first I was like ugggh what's this set to. Triple checked you had it in there and hit the ground running. Great to see that the input spec issue (when done properly) is a problem of yesteryear, not even a thought now!

 

[bikescene]

Thanks for the captures, made it easy to quickly test it all out, gave it a quick 10-15 min spin on the Red Mesa

MB DRMW Red Mdn Test STD
I took this as the standard method. Seemed to have a touch more low mids

MB DRMW Red Mdn Test xSTD
Same as the STD but with the tiniest shift from less low mids to a treble kiss, makes palm mutes a bit tighter

MB DRMW Red Mdn Test NewInput xSTD
Another notch in the treble direction over the one above. Somehow feels a touch louder or gainier than the others but I'm talking 0.5%, probably just an EQ thing

From a players perspective just sounds/feels like different options or someone nudged the EQ by 0.2 in the mids or treble. Would be interesting to see how the new stuff NULLS against the original amp. Either way it's been a good quick test just to see how it plays first hand. My initial assessment is same same but mildly different, not world changing at all. If there's any benefit to all of this it's probably on the mixing side of things.

As a side note the embedded input spec is great. At first I was like ugggh what's this set to. Triple checked you had it in there and hit the ground running. Great to see that the input spec issue (when done properly) is a problem of yesteryear, not even a thought now!

Thanks for checking out. The raw output with the wet/dry training seems to increase the output levels significantly in an unexpected way. I never checked if the Calibrated output fixed things. I’ve had to run all the Wet/Dry experiments with the Normalized Output. Good to know that the auto calibration tamed what probably was a real ugly sound.

I agree that the core tone is already in the normal training. To me, it also sounded like the couple of slight EQ moves that you mentioned. Electric guitar can easily be mangled by small EQ changes which makes the hair splitting insignificant.

It’s interesting to be able share sounds in this manner and see how others experience the nuances to the tone.

It’s cool that TONEZONE3000 subtlely opened things up like this. It’s there for those who want to tweak, but pretty much still invisible.

 

[[Nathan]]

Thanks for checking out. The raw output with the wet/dry training seems to increase the output levels significantly in an unexpected way. I never checked if the Calibrated output fixed things. I’ve had to run all the Wet/Dry experiments with the Normalized Output. Good to know that the auto calibration tamed what probably was a real ugly sound.

I agree that the core tone is already in the normal training. To me, it also sounded like the couple of slight EQ moves that you mentioned. Electric guitar can easily be mangled by small EQ changes which makes the hair splitting insignificant.

It’s interesting to be able share sounds in this manner and see how others experience the nuances to the tone.

It’s cool that TONEZONE3000 subtlely opened things up like this. It’s there for those who want to tweak, but pretty much still invisible.

Yeah, and I love the open source nature of NAM and that a community of people can hone in on certain aspects of it and try and come up with some novel solutions (some will be pointless, some will be negligible but others might be a big breakthrough).

Definitely would have been a time in my life where I'd be all over this stuff but I'm just a bit too busy these days. I'm still here for it but kind of just reaping the rewards off of everyone's findings. TZ and TH have definitely been massive helps here, what a crazy time for tech we all have access to.

 

[DLC86]

The raw output with the wet/dry training seems to increase the output levels significantly in an unexpected way.

Yep, that's an issue of that type of training on tonezone. I wrote to Woodbury (one of tonezone's admins) and he told me they were already aware of this issue and will fix it soon.

 

[DLC86]

MB DRMW Red Mdn Test STD
I took this as the standard method. Seemed to have a touch more low mids

MB DRMW Red Mdn Test xSTD
Same as the STD but with the tiniest shift from less low mids to a treble kiss, makes palm mutes a bit tighter

MB DRMW Red Mdn Test NewInput xSTD
Another notch in the treble direction over the one above. Somehow feels a touch louder or gainier than the others but I'm talking 0.5%, probably just an EQ thing

From a players perspective just sounds/feels like different options or someone nudged the EQ by 0.2 in the mids or treble. Would be interesting to see how the new stuff NULLS against the original amp. Either way it's been a good quick test just to see how it plays first hand. My initial assessment is same same but mildly different, not world changing at all. If there's any benefit to all of this it's probably on the mixing side of things.

This is the kind of difference I usually hear in the captures I've tested as well, but by analyzing them there was really no boost in the high frequencies, I think we get that impression cuz there's less aliasing and the high end is less cluttered... and by directly comparing it to what I was capturing, that sounded closer to the real deal.

 

[DLC86]

Guys, I'm really sorry but I realized now that I copied the wrong architecture code in my post (both here and in the facebook group), i picked up my88886 instead of my8886_16 from the core.py file... already fixed it in that post.
For anyone who copied it from there it's not a big deal though, they're pretty much identical in terms of accuracy, the wrong one just uses a higher number of parameters and thus more cpu on the final model.

 

[Humbug]

I had a little bit of time last night to reamp and applied some of the recent training info from this thread. I thought I'd share for anyone who finds it interesting.

I put a bunch of NAM files in the following folder:
https://drive.google.com/drive/folders/12ikGAk-5YSD1NhHzzDIdBO0mWHTIeark?usp=drive_link

I made captures of 4 amp channels:
-Mesa Dual Recto Multi-Watt Red Channel Vintage Mode
-Mesa Dual Recto Multi-Watt Red Channel Modern Mode
-Peavey 6505 1992 Original Lead Channel
-Vox AC15C1 Normal Channel

With the Wet/Dry method, I trained the files in the following ways:
-Standard
-xSTD code with default learning rate/decay
-My training file with some higher freq sweeps a la the 50K file + xSTD code as above

At minimum, there's 12 captures you can cycle through to hear how the different training is affected with my setup.

I had a chance to play around with the AC15 models today, in a sort of half assed blind A/B-test where I just played and switched between them to see if any of them felt better than the others. The "Newinput..." came out on top. Not by a huge margin, but significant enough for me to go "that one".

 

[staas]

Great to see all the discussion here! Yes, you’ve discovered that you can now add custom architectures during training on TONEZONE3000. This is highly experimental—we’re encouraging everyone to explore and push NAM forward. Since Woody and I have taken over maintenance of ToneHunt we've been focused on integrating ToneHunt and TONEZONE3000 into one site (launching soonish!). If you have questions, spot bugs, or need help, let us know :)

 

[HomespunEffects]

Hey, the site is working now! :)

 

[staas]

Hey, the site is working now! :)

Hey that's great to hear! I was just about to message you. This was a bug on our end that occurred during the account creation process. So sorry for the hassle. Thank you for reporting it!

 

[DLC86]

I managed to get close to the 50k input aliasing performance but with a much shorter file, which leads to much less training time... I'll do more tests to see if I can improve it further, but in the meantime here it is for you folks:

 

[2dor]

I managed to get close to the 50k input aliasing performance but with a much shorter file, which leads to much less training time... I'll do more tests to see if I can improve it further, but in the meantime here it is for you folks:

View attachment 37971

Nice. Does the loudness side-effect occur with this shorter signal?

 

[bikescene]

I managed to get close to the 50k input aliasing performance but with a much shorter file, which leads to much less training time... I'll do more tests to see if I can improve it further, but in the meantime here it is for you folks:

View attachment 37971

Nice! I will give it a try.

Nice. Does the loudness side-effect occur with this shorter signal?

The added volume seems to be a product of the wet/dry training. I tried the regular V3 file and the volume bump is still there. I think @DLC86 mentioned that he reported it to the TZ3000 group.

 

[DLC86]

Nice. Does the loudness side-effect occur with this shorter signal?

What @bikescene said, it's an issue on tonezone, they should fix it pretty soon

 

[Audio Transformer]

Now the 60K is by far the best. Anybody tried it? In the new Facebook group « Neural amp modeler LABS »

 

[laxu]

Now the 60K is by far the best. Anybody tried it? In the new Facebook group « Neural amp modeler LABS »

Timbuck3 waiting for the 100K, with a stereo fx loop.

 

[DLC86]

We've made quite some progress in the last few days, here's a graph comparing standard input and a couple previous versions with Francois's "3D" and my "TTSv10" inputs, which are currently the best performing ones. All are trained with the xSTD architecture and all coming from my VH4 pedal with the gain at noon (which is already a lot of distortion)

Basically we managed to lower the aliasing level by more than 20 dB on average (compared to standard) just by fine tuning the sine sweeps in the input file. And hopefully there's still some room for improvement.

You can download my TTSv10 input here:


PS: the graph is made using "Cliff's method", i.e. a -12 dBFS sine sweep ranging from 9 to 11 kHz which produces the 2nd harmonic at 18-22 kHz, all the rest you see is aliasing.

 

[Slammin Mofo]

Thank you so much for the amazing feedback! We're working on a big update that will include some of your requests. I handle design at TONEZONE3000, so I'll share your code with my co-founder, who manages the engineering.

Thanks for including a custom architecture option! Unfortunately, I'm still not able to use my hyper accuracy model in custom mode as it exceeds the allowed architecture size.

I totally understand this limitation, as it would take your servers almost triple the amount of time of a single standard capture to generate a single hyper accuracy capture, but it would be nice to have an option where you can purchase extra server training time for the more complex models.

In any case, it's a great product as it is even if the bigger architectures remain forbidden on it.

 

[Slammin Mofo]

We've made quite some progress in the last few days, here's a graph comparing standard input and a couple previous versions with Francois's "3D" and my "TTSv10" inputs, which are currently the best performing ones. All are trained with the xSTD architecture and all coming from my VH4 pedal with the gain at noon (which is already a lot of distortion)

Basically we managed to lower the aliasing level by more than 20 dB on average (compared to standard) just by fine tuning the sine sweeps in the input file. And hopefully there's still some room for improvement.

You can download my TTSv10 input here:


PS: the graph is made using "Cliff's method", i.e. a -12 dBFS sine sweep ranging from 9 to 11 kHz which produces the 2nd harmonic at 18-22 kHz, all the rest you see is aliasing.

Thank you for this low-aliasing training file! I already recorded clips with it and I can train them on TZ3000 without any problems via the dry/wet option, but if I want to train the output clips made with your TTSv10 input file, the local trainer (version 0.11.0) tells me the following:

"Input provided isn't silent for at least 19280 samples before the starting index. Responses to this non-silent input may leak into the dataset!"

As far as I can tell, your training signal is the standard V3 test signal with an added sine wave section at different amplitudes and the same two closing sections as the V3 files, the only difference is in length basically, so I don't understand why the "input is silent before starting index" is even a problem, as it follows the V3 file closely in the beginning and the end. Maybe I'm misunderstanding the input index terminology in this case.

What do I need to adjust so that the local trainer accepts your TTSv10 input file?

 

[DLC86]

Thank you for this low-aliasing training file! I already recorded clips with it and I can train them on TZ3000 without any problems via the dry/wet option, but if I want to train the output clips made with your TTSv10 input file, the local trainer (version 0.11.0) tells me the following:

"Input provided isn't silent for at least 19280 samples before the starting index. Responses to this non-silent input may leak into the dataset!"

As far as I can tell, your training signal is the standard V3 test signal with an added sine wave section at different amplitudes and the same two closing sections as the V3 files, the only difference is in length basically, so I don't understand why the "input is silent before starting index" is even a problem, as it follows the V3 file closely in the beginning and the end. Maybe I'm misunderstanding the input index terminology in this case.

What do I need to adjust so that the local trainer accepts your TTSv10 input file?

Honestly, I don't know
I've tried too a bit to make it work locally but haven't found where to tell the trainer to disable the checks on the input file. But the message you get is a bit strange indeed, in my case it told me it was the wrong input file.

The full version of the trainer should allow any type of inputs but the instructions are not clear on where to put the config files, so I gave up on that too.