NeuralDSP / Darkglass helix type thing....? Now with Rabea demo!

WiresDreamDisasters

lysander said:

I disagree, a neural network is function approximator, that has to be calibrated to the function it is trying to approximate. 

Given even the number of controls differs between different amps, it is very unlikely that there is a single trained network that would somehow work for every amp without need for retraining, I can’t think of a single comparable example in other fields.
Much more likely is that each amp model is a separate trained network, and a new amp means a new training.

And no, the SHARC application domain is completely different from a GPU.
DSP processors are primarily designed for low latency flow processing with little no parallelism and where the primary design constraint is real time operation.
GPUs are designed for extremely high parallelism of fairly primitive computation units, with very little consideration for latency.
The application domain is at the opposite end of the spectrum, there’s a reason no one does real time audio on GPU and people don’t use DSP chips for machine learning.

Okay so I've done some neural network stuff in Python a fair bit now. Classifying different types of guitar string articulation - whether something is a palm mute, a regular pluck, a pick-scrape, etc.

And once I've setup a model with training and testing data, I can run a prediction on completely new data that the model has never seen before, and it can predict with some accuracy the type of articulation the file is. Back propagation is not the most intensive part of the process, it's actually one of the fastest - https://www.guru99.com/backpropogation-neural-network.html#3

On DSPs, it depends what sort you're talking about:

SHARC chips are different to FPGA chips. But they're both technical DSP specialist chips because they can operate in near realtime. There are some frameworks like CUDA which allow you to pass off DSP related tasks to a GPU, so it's not really all cut and dry IMHO.

But yes generally... DSP > Audio .... GPU > Visual....

rossyamaha

I'd love to know what the fuck you're all talking about. 

fretmeister

rossyamaha said:

I'd love to know what the fuck you're all talking about. 

I don’t. I’m not a virgin. 

MikeB

Ola England will have the issue nailed with his YouTube series, "Will it chug" 

mixolyd

Doug Castro of Neural is commenting on a couple of threads on TGP.

***************

“ As for our modeling styles, Neural Capture simply learns the sound of the rig with the settings you have.

But the virtual devices we include with the product are as accurate models of the original circuits as modern science allows, that's what propelled us with the plugins, and why not stick to the one that brought you to the dance.”

****************

” Udo Solzer and Felix Eichas have some very interesting papers on approaches slightly closer (and potentially better) thank Kemper's. 

But our approach relies on a much more fluid structure. 

Kemper seems to follow a more rigid Weiner-Hammerstein topology, for which the non-linearity transfer function is static (hyperbolic tangent or something of the sorts) and the profiling process adjusting the FIRs and saturation block gain.

Our deep learning approach allows us to use more sophisticated loss functions (ours is based on human perception, which results in a much more "accurate" black-box model). Of course, you need a lot more computational power, which is another reason we had to make such a powerful unit: when using Neural Capture the neural network uses 4 SHARC cores in tandem for training).

This approach generalizes better (it can cope with a wider range of sounds), and can even learn parallel signal paths (such as multiband distortion or pedals with a clean blend, dual microphones).”

************

mixolyd

So it sounds like:

• profiling can be done in the unit, using their deep learning system

• the included models are not profiled in this way, allowing proper amp-like behaviour re EQ etc

• they believe their profiling is more sophisticated than that used by Kemper.

lysander

@WiresDreamDisasters classification which is what you’re doing is a different problem than modelling a continuous system and of course it still works out of sample without retraining otherwise it would be pointless.
But the fact it recognises things you haven’t trained it on is the equivalent to saying that an amp modelled by a NN still sounds right when you play into it something that it wasn’t trained on.
It’s not equivalent to a single NN being able to model several amps.

@fretmeister yeah yeah anyone with some technical knowledge and intellectual curiosity is a virgin, that sounds a bit insecure to be honest.

fretmeister

It’s lysander said:

@WiresDreamDisasters classification which is what you’re doing is a different problem than modelling a continuous system and of course it still works out of sample without retraining otherwise it would be pointless.
But the fact it recognises things you haven’t trained it on is the equivalent to saying that an amp modelled by a NN still sounds right when you play into it something that it wasn’t trained on.
It’s not equivalent to a single NN being able to model several amps.

@fretmeister yeah yeah anyone with some technical knowledge and intellectual curiosity is a virgin, that sounds a bit insecure to be honest.

It’s comedy.

Your reaction is the insecure bit.

Jonathanthomas83

I dont get it, so they have models in there...like the plugins, but allow profiling too? Confused now.

WiresDreamDisasters

lysander said:

@WiresDreamDisasters classification which is what you’re doing is a different problem than modelling a continuous system and of course it still works out of sample without retraining otherwise it would be pointless.
But the fact it recognises things you haven’t trained it on is the equivalent to saying that an amp modelled by a NN still sounds right when you play into it something that it wasn’t trained on.
It’s not equivalent to a single NN being able to model several amps.

@fretmeister yeah yeah anyone with some technical knowledge and intellectual curiosity is a virgin, that sounds a bit insecure to be honest.

Nah see this is what I think the Kemper is doing. It's got a selection of schematics and a selection of transfer function points where it selects from a list of the closest approximation to the input source. All you really need to achieve this stuff is to be able to say "the amp reacts in this way in these spots" and select the right functions for those parts of the schematic... then a final tone match to get it closer... and there you go.

You don't need to model a continuous system. You just need to break it up into constituent parts.

IMHO.

IE: We're not talking about a single NN. We're talking about several chained together that are listening for completely different things. Some are going to be listening for harmonic relationships whereas others are going to be listening for dynamic range. For example.

Jota

sgosden

That interface is just helix ... I wonder if L6 have any ownership / IP on that particular style. 
Or if it's even worth them challenging 

TTBZ

Surprised there have been no audio demos yet. The cab/IR section looks like a more intuitive way of doing it than anything else on the market.

TeleMaster

I don't why people are rushing to order without sound clips. If this sounds similar to the Helix amp models I'll probably pass. 

WiresDreamDisasters

I wont order it until I see clips of it doing it's neural capture thing with comparisons against the real amps.

Basically I was going to get a Kemper Stage this year, but if this thing can capture amps and run multiples of them at the same time, and have the same tone and dynamics and all that, then I'll be very interested.

sweepy

TeleMaster said:

I don't why people are rushing to order without sound clips. If this sounds similar to the Helix amp models I'll probably pass. 

If it sounds like their plugins then it’ll be a factor up from the Helix, it’s not as if it’s a product coming from a startup with no track record of serious audio design 

lysander

I don’t know, I think a large part of what makes people think their plugins sound good is the very shimmery reverb and delay they include that are the type that make anything sound huge.
But I already have good delay and reverb plugins and don’t need that.

I have their Plini plugin and never use it - I find the included cabs dreadful, and the amp themselves don’t give me anything I can’t dial in quicker in Helix, IMO.

fretmeister

WiresDreamDisasters said:

I wont order it until I see clips of it doing it's neural capture thing with comparisons against the real amps.

Basically I was going to get a Kemper Stage this year, but if this thing can capture amps and run multiples of them at the same time, and have the same tone and dynamics and all that, then I'll be very interested.

There's a live Q&A with Doug tomorrow...

https://www.youtube.com/watch?v=sqLZMl5zIn4

WiresDreamDisasters

lysander said:

I don’t know, I think a large part of what makes people think their plugins sound good is the very shimmery reverb and delay they include that are the type that make anything sound huge.
But I already have good delay and reverb plugins and don’t need that.

I have their Plini plugin and never use it - I find the included cabs dreadful, and the amp themselves don’t give me anything I can’t dial in quicker in Helix, IMO.

TBH I have two of their plugins, Fortin Nameless and Fortin NTS. 

I like both, but I prefer Nameless. And out of all of the amp modelling plugins I own, I think it sounds great!

I did a run through last year:


Fortin NTS is decent. But I find it a bit more confusing to use than Nameless, not sure why.

But these days I'm kinda done running amp-sims in the box. I just want to record pure audio with as little latency as possible. So I'd rather have a piece of outboard.

I use my Helix at home right now because all of my amps are at the studio and with Ownhammer Diezel cab impulses and the new Revv Generator Red model, I can get really good high-gain tones that way, so that's what I've been doing.

TeleMaster

WiresDreamDisasters said:

I wont order it until I see clips of it doing it's neural capture thing with comparisons against the real amps.

Basically I was going to get a Kemper Stage this year, but if this thing can capture amps and run multiples of them at the same time, and have the same tone and dynamics and all that, then I'll be very interested.

I wouldn't bother. I've had 4 of them and now the 4th one has been sent back to Kemper for replacement of the switches. Once they've confirmed the issue we've agreed that I'll get a full refund under my consumer rights. I absolutely love a lot about the Stage and Kemper overall, but I've had enough of the quality control. 

I've got a Revival Drive running into a SD Powerstage 170 with a HX Effects for..  Effects and it sounds awesome. The Kemper still had a bit of a digital sound to it when you upped the gain on the profile.