TONE3000 (previously ToneHunt and TONEZONE3000)

[2dor]

10% upon the standard architecture. If standard is 100%, xSTD is 110%

EDIT: it's actually a bit less than 10%
View attachment 37391

That xSTD architecture is cool if it's the one labeled MY8886_16 in his core.py file.

version_12 = reamp trained with the xSTD (MY8886_16) architecture
version_13 = same reamp trained with the vanilla Standard architecture

A couple of conclusions I can draw:
- training takes a bit longer for 1000 epochs but not crazy much (~8 minutes on my RTX 4070, AMD 5700X setup)
- high-end seems to be more faithful to the source (lower pre-emphasized MRSTFT) which is pretty cool
- ESR shows improvements too which is nice as well; shows that most of the spectrum (aside from the high-end) is faithfully represented
- The Standard vanilla architecture shows about 14% - 15% RT-CPU in Reaper on my machine
- The xSTD architecture shows a slight bump to 17% - 18% RT-CPU in Reaper on my machine

Overall, I think this new architecture combined with the enhanced "anti aliasing" sequences can help models behave better as far as the overall training is concerned (high-end and overall time-domain).

 

[DLC86]

That xSTD architecture is cool if it's the one labeled MY8886_16 in his core.py file.

Yes, that's the one

 

[2dor]

Yes, that's the one

Cool.
Percentage-wise, CPU demand increased by about 20% compared to the vanilla Standard architecture. Still manageable on modern machines.

 

[Mike Oliphant]

The nice thing about this is that it's an almost "free" improvement cpu wise, the super input itself doesn't produce heavier models,

Keep in mind that by adding extra high-frequency data to the training signal, you are likely having some impact on the model's ability to match other (arguably more important) frequencies.

Given how accurate the beefier NAM models are, this may well be a good tradeoff. It is something to be aware of, though.

 

[HomespunEffects]

So what do I have to do to get my models after I upload my stems?

 

[2dor]

So what do I have to do to get my models after I upload my stems?

Just click on "Your Tones" and you should be able to find the profiles there.

The ones with the Green checkbox have finished training successfully; click on any of them and then on Download.

 

[HomespunEffects]

Ok. Thanks. I don’t get that menu bar on the left though. Do i have to use a special browser? Safari doesn’t work with the site at all except the home screen. Brave works with the site but i guess not all of it.

If I can’t work a web browser, maybe tonezone3000 isn’t for me.

 

[laxu]

Ok. Thanks. I don’t get that menu bar on the left though. Do i have to use a special browser? Safari doesn’t work with the site at all except the home screen. Brave works with the site but i guess not all of it.

If I can’t work a web browser, maybe tonezone3000 isn’t for me.

Safari seems to work totally fine. That sidebar is what you see when you click on your user profile at the top right.

 

[HomespunEffects]

Safari seems to work totally fine. That sidebar is what you see when you click on your user profile at the top right.

I’ll get a screen shot.

 

[HomespunEffects]

Home page on Safari:

After clicking account avatar:

Home page is the same on Brave but after clicking the account avatar I get this:

Attempting to view source, I get "Too many redirects" error.

 

[2dor]

Home page on Safari:
View attachment 37440
After clicking account avatar:
View attachment 37441
Home page is the same on Brave but after clicking the account avatar I get this:
View attachment 37442

Attempting to view source, I get "Too many redirects" error.

Looks like some issues on the server-side? If you try an "incognito" window, does it behave the same?

 

[laxu]

Home page on Safari:
View attachment 37440
After clicking account avatar:
View attachment 37441
Home page is the same on Brave but after clicking the account avatar I get this:
View attachment 37442

Attempting to view source, I get "Too many redirects" error.

I don't see any such issue which points to this likely being a problem on your local machine, and it's just unable to load something. I would try clearing your browser cache and cookies for this site, disabling extensions, and if you have whatever VPN is there, that too.

 

[HomespunEffects]

I don't see any such issue which points to this likely being a problem on your local machine, and it's just unable to load something. I would try clearing your browser cache and cookies for this site, disabling extensions, and if you have whatever VPN is there, that too.

Brand new Mac mini. Brave was installed just for this. No VPN. I’m not really worried about getting this working. Thanks though.

 

[DLC86]

The anti-aliasing just got even better!
Here's a comparison done on one of my captures (Cornish G2 clone)

Standard

"50k" input

"50k" input + xSTD architecture

Comparison of the three with Cliff's method

And here's an audio comparison:

Standard


50k


50k + xSTD


Basically in the last one the aliasing is unhearable even with a 0 dBFS sweep, only slightly hearable when the fundamental gets close to 24 kHz (something that will never happen with a guitar signal), I'm impressed!

The only downside of this is that the training takes quite a bit longer than standard, but it's worth it imho, especially considering that the impact on the cpu is minimal in the case of the xSTD architecture and absent with the "50k" input alone.

 

[[Nathan]]

I'm not upto scratch in all the anti aliasing stuff.... has anyone compared the original training file/process with this new one? Do the models sound or respond better (I'm talking actual amps not sweeps).

 

[DLC86]

I'm not upto scratch in all the anti aliasing stuff.... has anyone compared the original training file/process with this new one? Do the models sound or respond better (I'm talking actual amps not sweeps).

Yes, to my ears they're slightly better and also null tests are a bit better... I'll probably make a video about this soon

 

[2dor]

I'm not upto scratch in all the anti aliasing stuff.... has anyone compared the original training file/process with this new one? Do the models sound or respond better (I'm talking actual amps not sweeps).

Yeah, I did take a stab (more than 1 actually, quite a few literally) at both implementing the extra sweeps into the training signal as well as training with the xSTD architecture.

My conclusion is, playing guitar through the "anti-aliasing" + xSTD trained model and the STANDARD/VANILLA one, I can't really tell a difference.

That said, by the looks of those graphs, it could be a great way to train models if you want to use them in large mixing sessions (I recall some folks who do producer's work complain about aliasing piling up between multiple instances to the point where it's really tough to do anything about the muddied up results).

If I ever train anything new that I want to share with the world, I'll probably use these findings but I still think most folks aren't missing out even if they're using the STANDARD, vanilla models.

 

[DLC86]

Here is my model if you wanna try it, both standard and "50k xSTD":

TTS Cornish G2 (calibrated + anti-aliasing) | ToneHunt

Pete Cornish G-2 (Aion FX clone) - Clone of the Pete Cornish G-2 based on the schematic traced by Aion FX It has four D9E diodes with a Vf comprised between 0.21 and 0.23 2 files - "standard" was captured with v3.0.0 input file and standard architecture - "50k xSTD" is the anti-aliasing version...

tonehunt.org

 

[BenIfin]

Yes, to my ears they're slightly better and also null tests are a bit better... I'll probably make a video about this soon

Hey Paisano ! :)

This is all genuinely beyond my actual understanding ..... that said ..... is it correct to say that the aliasing in a [current gen] Standard NAM Capture and a Standard Tonex Capture is essentially the same ?

Also from my reading of the above, and previous Fractal tests that have been posted around the traps, this new NAM approach seems to me (?) to produce substantially less aliasing than even an Axe 3 (?)

It all kind of begs the question, why didn't/haven't IK and Steve A. "updated" their Training and Capturing to incorporate something like the above ? Surely they have known about this aliasing issue since Day 1 ?

All the best,
Ben

 

[DLC86]

This is all genuinely beyond my actual understanding ..... that said ..... is it correct to say that the aliasing in a [current gen] Standard NAM Capture and a Standard Tonex Capture is essentially the same ?

Nope, tonex is slightly worse than nam cuz it runs at 44.1 kHz internally (the plugin, haven't tested the pedals but probably the same applies).

And it is much worse when it runs at a sample rate different than 44.1 kHz cuz it applies a shitty SRC. Here's what happens at 48 kHz:

Also from my reading of the above, and previous Fractal tests that have been posted around the traps, this new NAM approach seems to me (?) to produce substantially less aliasing than even an Axe 3 (?)

It looks quite good but no one has compared it properly to other devices yet, so that remains to be seen... don't trust those statements you read on the NAM group
Imho it's not that easy to beat the axe fx III cuz it has 32x oversampling iirc.

It all kind of begs the question, why didn't/haven't IK and Steve A. "updated" their Training and Capturing to incorporate something like the above ? Surely they have known about this aliasing issue since Day 1 ?

You should ask Steve about that