Ten billion profiles/captures = one amp

[MirrorProfiles]

Yeah I've posted the article a few times I think. But where are the results?

Since the TINA reveal we have seen like only a few new models. You would think this would result in them updating or adding models at a fairly rapid pace.

There’s probably no benefit to them dumping loads of new plugins out if they aren’t selling well. Frustrating watching them drip feed, but Mayer and maybe Misha’s aside, there seems to have been steadily less interest in each new release.

 

[Lysander]

We truly don't know how long TINA has been used.

It was announced with a fanfare back in July 2024 thou.

 

[Mongillo19]

It was announced with a fanfare back in July 2024 thou.

Right but Nameless X was a few months before that and Fortin has explicitly said TINA was used on it

 

[Richard_G]

Fortin has stated that TINA has been used on more than the Nameless as well, like in this post on Instagram in February:

 

[Orvillain]

There's much more to making an amp and releasing it, than the actual content generation steps. There's a fuck ton of marketing, sales, infrastructure stuff that needs to happen.

I can't be arsed to run the numbers. But... it is entirely possible that NDSP are going to drop a shit ton of amps in the future in a NDSP Native type plugin... which I'd be totally down for.

The truth is, beyond a certain point, we're all just guessing.

Sorry, I'm in NDSP fanboi mode right now!

 

[KHAAAAAAANNNNNNN!!!!!!!!]

I don't buy it simply because of the logistics.

For example, let's consider a relatively simply amp with six knobs: Gain, Bass, Mid, Treble, Presence, Master Volume. Now let's say we use a fairly coarse resolution and sample each knob at only 10 positions (1, 2, 3, ..., 10).

That's 1M possible combinations. 1 million!!!

Now let's assume that this so-called machine learning can learn the response in a mere 120 seconds which is extremely optimistic. That means it will take 120M seconds to learn all possible combinations of the controls.

120M seconds = 2M minutes = 33,333 hours = 1,388 days = 3.8 years.

Add a Depth knob and the time increases to 38 years.

Don't forget the bright switch...

 

[Orvillain]

 

[BenIfin]

... Sorry, I'm in NDSP fanboi mode right now!

Its ok. Admitting the problem is the first step

 

[Piing]

I don't buy it simply because of the logistics.

For example, let's consider a relatively simply amp with six knobs: Gain, Bass, Mid, Treble, Presence, Master Volume. Now let's say we use a fairly coarse resolution and sample each knob at only 10 positions (1, 2, 3, ..., 10).

That's 1M possible combinations. 1 million!!!

Now let's assume that this so-called machine learning can learn the response in a mere 120 seconds which is extremely optimistic. That means it will take 120M seconds to learn all possible combinations of the controls.

120M seconds = 2M minutes = 33,333 hours = 1,388 days = 3.8 years.

Add a Depth knob and the time increases to 38 years.

You forgot to add the reaction of the amplifier when rolling down the guitar volume knob. A feature that, unlike the profiles/captures, is faithfully modeled at the Fractal units.

Just another exponential....

 

[Orvillain]

unlike the captures

This is not an accurate statement.

 

[Piing]

This is not an accurate statement.

If that is the case, I must say that my ToneX doesn't get it right. I don't know about other profilers because I haven't spent much time with them.

 

[Orvillain]

If that is the case, I must say that my ToneX doesn't get it right. I don't know about other profilers because I haven't spent much time with them.

If you read the paper posted above, it is clear that the parametric approach does capture the behaviour of the reference amp.

However what you as a user are creating with ToneX, NAM, QC, etc... are not parametric models. They are snapshot models, and will of course have limitations as you've experienced.

 

[yeky83]

I'm bumping this thread.

End-to-End Amp Modeling: From Data to Controllable Guitar Amplifier Models

I can't remember if we discussed this already or not. So this might be old news, but that paper is written by the NDSP team.

• They train a neural network to model an amp end-to-end from audio data
• The model takes input audio + knob settings as inputs
• Knob settings are randomly sampled across the full control range
• This randomisation helps the model learn interactions between controls, not just isolated positions
• The result is a fully continuous, controllable amp model
• Training data comes from real amps with automated, randomised parameter capture
• No schematic or circuit knowledge is required (pure black-box approach)
• They use an LSTM to capture nonlinear + time-dependent behaviour

Old news:

Post in thread 'NDSP Quad Cortex'

Sep 24, 2024

You can read about NeuralDSP's approach here: https://arxiv.org/html/2403.08559v1

They don't use just X fixed values for each knob but a randomized sample. While this might mean that the model is not accurate at every single possible permutation, in practice it's probably in that "nobody cares" territory where you can't really notice it.

• laxu

• 

If you read the paper posted above, it is clear that the parametric approach does capture the behaviour of the reference amp.

However what you as a user are creating with ToneX, NAM, QC, etc... are not parametric models. They are snapshot models, and will of course have limitations as you've experienced.

Has nothing to do with his complaint which was how the model reacts to his guitar volume knob. “Snapshot models” can do that well too.

 

[laxu]

There’s probably no benefit to them dumping loads of new plugins out if they aren’t selling well. Frustrating watching them drip feed, but Mayer and maybe Misha’s aside, there seems to have been steadily less interest in each new release.

You'd still think there would be more new amps, or upgraded amps for the QC. When they released a few Fenders I expected they'd slowly update all the amps to the new modeling but haven't done so for some reason. Either that means their approach has unsolved issues or that it's much slower to do than expected.

 

[MirrorProfiles]

You'd still think there would be more new amps, or upgraded amps for the QC. When they released a few Fenders I expected they'd slowly update all the amps to the new modeling but haven't done so for some reason. Either that means their approach has unsolved issues or that it's much slower to do than expected.

Aside from all the other work involved in getting them to something that can be released, there’s probably little benefit in blowing their load constantly.

 

[Orvillain]

Old news:

Cool, thanks for letting me know.

Has nothing to do with his complaint which was how the model reacts to his guitar volume knob. “Snapshot models” can do that well too.

Cool, thanks for letting me know. I appreciate the education.

 

[laxu]

Aside from all the other work involved in getting them to something that can be released, there’s probably little benefit in blowing their load constantly.

Which is why I mentioned slowly. Like let's say every major update they overhauled 2-3 amp models, same as say Line6 does for Agoura. People would be pretty happy about that, right?

But for the past year, excluding the plugin compatibility updates, there's been a whopping 3 amp models released afaik.

 

[MirrorProfiles]

Which is why I mentioned slowly. Like let's say every major update they overhauled 2-3 amp models, same as say Line6 does for Agoura. People would be pretty happy about that, right?

But for the past year, excluding the plugin compatibility updates, there's been a whopping 3 amp models released afaik.

Who knows how much that actually leads to more sales of HW. Presumably there’s enough wind in their sales with all the Neural Capture V2 and John Mayer affiliations to not need to do much.

It could also be that they don’t think any existing QC models need overhauling, in their mind, or that the benefits don’t make sense until later FW changes.

I’d imagine releasing new models as plugins with eventual PCOM is their most financially appealing way of doing it. And dumping those too frequently doesn’t generate as much income or HYPE as staggering things out.

 

[Jarick]

I’d be curious what their sales look like nowadays. The Mayer plugin was huge because it’s Mayer but outside that there’s not been much fanfare around new plugins.

If I were a QC owner I’d be more interested in new effect models because in theory you could just get more amps via captures.

 

[GTR37]

Which is why I mentioned slowly. Like let's say every major update they overhauled 2-3 amp models, same as say Line6 does for Agoura. People would be pretty happy about that, right?

But for the past year, excluding the plugin compatibility updates, there's been a whopping 3 amp models released afaik.

I think they are honestly just backlogged in PCOM
I don’t know how big a design team they have but the had 2 new products , some significant updates to Nano
I think it will be interesting to see where the next update goes I would hope a couple of new amps but as for effects what are people still needing
The Delays / Reverbs /Mods are really good now
Maybe some drive boxes , feedbacker ,
New cabs
I honestly can’t think of much it would be cool if they did a couple of less common amps like a Mezzabarba , Driftwood , Badlander
Jose etc