Neural DSP Users?

[RexRemus]

Video embargo lifted today - so I think we'll all see a lot more info dropping soon and a lot less speculation.

I disagree to some extent that I definitely see the use case of an ML approach - that doesn't instantly make me think it's "best", only that I understand the thought process of addressing it as an ML challenge because if you're modeling the complexities of a biological organism (us) and their hearing, then there are opportunities to optimize that in certain ways via a trained model vs a procedural algo. They STILL need to make a good model, but there's nothing inherently "wrong" about tackling the problem or capture in that way.

[Top-L]

On the "AI" thing - it's absolutely marketing but they do it because in 99% of people's heads machine learning (which is what they are doing) and "AI" are the same thing - and it's not really their fault. Companies like Google, Apple, Tesla, and many others have all dubbed machine learning as "Ai" because AI sells basically.

So yes Neural isn't building generalized artificial intelligence, but they have built and trained a very specific and sophisticated ML algorithm to do their capture. While what they are doing exactly will obviously never be talked about, they do talk about it hearing "like a human" and so I suspect they have invested a lot of time into trying to train a model that is aware of (and hence captures) the psychoacoustic anomalies of the human auditory system. Which is really quite interesting. If you've done music and mixing for any length of time you know there are just as many auditory "illusions" as there are visual ones - our brains and ears play tricks on us in certain ways. That's what led to the loudness wars for a while - because "louder" sounded "better" - only it didn't really. It's the basis for a lot of compression technology because frequency masking is a thing and the "loudest" thing at any given moment masks a quieter thing in the same frequency ranges so we can preserve the loud thing, discard info for the quieter thing and our dumb human hearing can't really tell the difference. We are more sensitive to 5-7k (baby cry) because.. we just are, cause babies crying, and so on - we perceive certain frequency balances more or less depending overall volume, etc, etc

So, I believe neural have spent time trying to model THAT. Not to take a flawless, 20-20k perfectly flat "sample" and determine the precise deltas between source and response. I think their capture tries to "hear" the response as a human ear would, with all of its dumb little quirks and illusions. Without that, yes you can capture a perfect "tonal" profile - you can reverse engineer an eq curve and determine how a very specific waveform you fed in gets clipped, you can do Fourier analysis to determine harmonics and overtones... but sometimes there's just a weird... swell, or a "bubble" or a grunt, or thump, or any other esoteric term we try to use to describe the "feel" of an amp. but we KNOW it's there we HEAR it - like a human. And they happen over time, things bloom or diminish over time, so you need to sample and "learn" the response of what you're capturing a certain way to get that. I think Neural - whatever they are doing - has THAT, a little something extra. Something that doesn't just get very close on tone, but also gets more (not all) of the "feel" because it's "hearing" things like people do. It's not a rigid algorithm, it's adaptive, it's "learning" (so yes, misnomered as 'intelligent') and they can keep training it on new things, making it better, and then update the models on all the QCs. It's the adaptivity, I think, that sets their POV on how to capture apart from the rest.

So yes... "AI" is marketing. but it's not fair to discount their approach to the technology of capture as garbage or just more of the same because they use a marketing term they basically MUST use because others before them have abused it so much that it's the only thing Joe Public will understand and respond to when referring to machine learning and adaptive algorithms.

And this isn't meant to sway you, just wanted to clear up the marketing nonsense POV - yes it's hype, but it's also unique and has a very real basis in some good technology.

Disclosure - I work in technology for a silicon valley company and actively build, train, and use ML/AI models for various parts of my day job. And yes, much of what I said is speculation because Neural isn't going to give up the special sauce, but I at least have a basis in understanding how and what I think they are trying to do and I suspect I'm in the ballpark.

I am currently getting my Master's in Computer Science (undergrad CS too), am actually taking a KR&R course atm; haven't taken ML yet. That said, I can't think of any reason/benefit to apply ML to performing a capture, which is something that can be done procedurally (Kemper).

Sometimes when you have a hammer, everything looks like a nail. One of their developers may be trained in ML, but that doesn't mean there is any tangible benefit from using it in captures, beyond marketing.

Unless they explain the algorithm (they won't), my gut reaction is that its ALL marketing. There may be elements of Machine Learning in their code, but the problem of profiling an amp doesn't suggest machine learning. Maybe, it will do something unique, like let the user twist all the eq and gain knobs during the profiling process to generate an intelligent profile that encompasses EQ band/sweep, multiple guitars, etc so the profile actually works like the real amp. That would appeal to me.

These auto tuning and profiling solutions have been used in car audio for 20+ years. They may be taking a different (not necessarily more effective) approach to avoid patent/copyright infringement. IDK the legal situation around Kemper, but if Kemper holds patents or copyrights on using auto tune for guitar rigs, if Neural all of a sudden has the same technology it could cause issues (not lawyer), they may be trying to skirt any legal issues (again, not lawyer.)

On TGP today there was a lengthy post by someone who recognized the same thing I did; that all of the demos sound the same. Earlier in this thread I mentioned it sounds like they don't have many models yet, but it may just be the Neural "sound". I will have to download their plugin demos at some point. They sound good but I'm not about to spend $200 on a plugin with a few artist sounds, so better to not even try them. Whole business model looks like took a page from Apple. I'm a PC/Linux/Android guy.

[RexRemus]

On the "AI" thing - it's absolutely marketing but they do it because in 99% of people's heads machine learning (which is what they are doing) and "AI" are the same thing - and it's not really their fault. Companies like Google, Apple, Tesla, and many others have all dubbed machine learning as "Ai" because AI sells basically.

So yes Neural isn't building generalized artificial intelligence, but they have built and trained a very specific and sophisticated ML algorithm to do their capture. While what they are doing exactly will obviously never be talked about, they do talk about it hearing "like a human" and so I suspect they have invested a lot of time into trying to train a model that is aware of (and hence captures) the psychoacoustic anomalies of the human auditory system. Which is really quite interesting. If you've done music and mixing for any length of time you know there are just as many auditory "illusions" as there are visual ones - our brains and ears play tricks on us in certain ways. That's what led to the loudness wars for a while - because "louder" sounded "better" - only it didn't really. It's the basis for a lot of compression technology because frequency masking is a thing and the "loudest" thing at any given moment masks a quieter thing in the same frequency ranges so we can preserve the loud thing, discard info for the quieter thing and our dumb human hearing can't really tell the difference. We are more sensitive to 5-7k (baby cry) because.. we just are, cause babies crying, and so on - we perceive certain frequency balances more or less depending overall volume, etc, etc

So, I believe neural have spent time trying to model THAT. Not to take a flawless, 20-20k perfectly flat "sample" and determine the precise deltas between source and response. I think their capture tries to "hear" the response as a human ear would, with all of its dumb little quirks and illusions. Without that, yes you can capture a perfect "tonal" profile - you can reverse engineer an eq curve and determine how a very specific waveform you fed in gets clipped, you can do Fourier analysis to determine harmonics and overtones... but sometimes there's just a weird... swell, or a "bubble" or a grunt, or thump, or any other esoteric term we try to use to describe the "feel" of an amp. but we KNOW it's there we HEAR it - like a human. And they happen over time, things bloom or diminish over time, so you need to sample and "learn" the response of what you're capturing a certain way to get that. I think Neural - whatever they are doing - has THAT, a little something extra. Something that doesn't just get very close on tone, but also gets more (not all) of the "feel" because it's "hearing" things like people do. It's not a rigid algorithm, it's adaptive, it's "learning" (so yes, misnomered as 'intelligent') and they can keep training it on new things, making it better, and then update the models on all the QCs. It's the adaptivity, I think, that sets their POV on how to capture apart from the rest.

So yes... "AI" is marketing. but it's not fair to discount their approach to the technology of capture as garbage or just more of the same because they use a marketing term they basically MUST use because others before them have abused it so much that it's the only thing Joe Public will understand and respond to when referring to machine learning and adaptive algorithms.

And this isn't meant to sway you, just wanted to clear up the marketing nonsense POV - yes it's hype, but it's also unique and has a very real basis in some good technology.

Disclosure - I work in technology for a silicon valley company and actively build, train, and use ML/AI models for various parts of my day job. And yes, much of what I said is speculation because Neural isn't going to give up the special sauce, but I at least have a basis in understanding how and what I think they are trying to do and I suspect I'm in the ballpark.

[Top-L]

My Fractal FM3 is the best feeling modeler I've tried, certainly over the AX8 and the Helix. I don't have much use for the USB recording outs of the FM3, although a recent update is said to fix it. Also, recent updates brought the firmware in line with the modeling (and models) or the Axe III.

My next processor will either be an FM3 or Gt1000.

[Mincer]

My Fractal FM3 is the best feeling modeler I've tried, certainly over the AX8 and the Helix. I don't have much use for the USB recording outs of the FM3, although a recent update is said to fix it. Also, recent updates brought the firmware in line with the modeling (and models) or the Axe III.

[Top-L]

Kemper broke new ground with the profiling tech, but that doesn't mean there isn't room for improvement. After all, a '90s Honda Civic will get you from point A to point B just fine, but a Tesla will do it faster, more efficiently, and more enjoyably. Similarly, borrowing from successful interfaces like the Helix and making improvements is a positive thing in my book.

I also really don't think that the QC absolutely has to do something entirely "new" and niche to be valid in the marketplace. Plus, it's beneficial for a product like this to have a lot of support up front to keep it evolving and become successful out of the gate. As I recall, the Kemper was subject to similar hype when it was emerging and that early adoption/interest paid off long-term.

On a side note, given Zoom's history and my own experience with their G5, I would rather spend $1600 on the QC than $800 for a flooboard that will be completely unsupported in another year or two, have a flakey fanbase, suffer terrible warranty support, and have horrible resale value to boot.

I think the issues most of us are contending with are that a) the QC is very expensive and b) we're reaching a point of diminishing returns in terms of what "new" technology can really offer the player.

As for the way the controls work on the QC, I believe you can do everything you mentioned by touch, if desired. That said, the foot/rotary switches have been tested vigorously to withstand significant abuse, so I doubt those will be the source of too many failures. However, it does seem a bit odd to be using your hands to turn knobs your dirty shoes have been smashing regularly.

I'm neutral on it. I wasn't particularly excited about the Kemper, the QC doesn't move the needle much for me. Although if I was a Kemper owner, if that was my approach, I would give it a serious look.

[Masta' C]

Kemper broke new ground with the profiling tech, but that doesn't mean there isn't room for improvement. After all, a '90s Honda Civic will get you from point A to point B just fine, but a Tesla will do it faster, more efficiently, and more enjoyably. Similarly, borrowing from successful interfaces like the Helix and making improvements is a positive thing in my book.

I also really don't think that the QC absolutely has to do something entirely "new" and niche to be valid in the marketplace. Plus, it's beneficial for a product like this to have a lot of support up front to keep it evolving and become successful out of the gate. As I recall, the Kemper was subject to similar hype when it was emerging and that early adoption/interest paid off long-term.

On a side note, given Zoom's history and my own experience with their G5, I would rather spend $1600 on the QC than $800 for a flooboard that will be completely unsupported in another year or two, have a flakey fanbase, suffer terrible warranty support, and have horrible resale value to boot.

I think the issues most of us are contending with are that a) the QC is very expensive and b) we're reaching a point of diminishing returns in terms of what "new" technology can really offer the player.

As for the way the controls work on the QC, I believe you can do everything you mentioned by touch, if desired. That said, the foot/rotary switches have been tested vigorously to withstand significant abuse, so I doubt those will be the source of too many failures. However, it does seem a bit odd to be using your hands to turn knobs your dirty shoes have been smashing regularly.

[Top-L]

And a quad-core processor that's far and above anything available in current modelers from what I gather.

I agree that tone and feel are "debatable", but the Andertons guys have demo'd just about everything and even they claim this is next level in terms of sound and feel, so that's intriguing at least.

Just to be clear, I'm not arguing. But if Neural can "build a better car", so to speak, I'm all for it.

The whole game is about marketing and excitement, imo. You've got products like Helix which to my ears has inferior modeling, but somehow has a large market share. Go figure. My impressions of what is available:

The Helix has a large presence on the forums, they spend a ton on advertizing and yet I personally think it sounds like last generation modeling. I did not like the Native demo, and I did a deep dive in the features. But some people love it, I suspect they love the interface and their ears are different than mine. I liked alot of the demo videos until I actually played with it.

The Fractal stuff... has the exclusivity of a custom shop.. yet in A/B demos I don't hear it being any better than the other ones. Yet there are people that will tell you its the best. Even though the FM3 is still plagued by hardware and software issues. (Does it yet function as a usb interface?) They have some highly skilled demo artists creating daily content for it. I enjoy these videos and think it sounds great.

Kemper. If you go to TGP you will read there are people who have bought and sold the Kemper multiple times. Like they can't make up their minds. It sounds great, but it involves countless hours hunting profiles. When you change the gain and eq its no longer like the real amp it profiled. I think there is some evidence that the whole profiling paradigm has some issues of accessibility for typical users who want their processor to work like a virtual collection of amps and pedals.

GT1000- Boss is like the Toyota Camry of the modeling world. Its not exciting, but when you look close it has a ton of really smart features. In AB testing it can be made to sound the same as the other ones. Some people will tell you its junk, not at the same level. Others tell you it feels better than the rest. I have many years with boss digital processors, I understand their paradigm and I know they can be made to sound good. Roland makes their own dsp so no one knows how it compares to ots sharc processors.

Headrush- Has solid modeling, Imo as good as it gets, but the interface is "form over function". Its pretty and easy to use, but the box is missing core features like amp switching. Its really huge, too big to travel with or take to work. The Gigboard is intriguing because its smaller, but features still kinda thin.

Zoom- The new box is brutalized for looking ugly (I dont agree), although its a modern incarnation of the much loved ME-80, in digital form. I think its a great idea, but I haven't heard any serious demos. I dont know how it sounds and for 800 clams, it should have amp switching. Id rather have amp switching than a color touch screen. I bet if I owned this I'd enjoy it.

The Quad Cortex is in the honeymoon marketing phase. Obviously there are tons of house-bound musicians who need something new to pass the time, so excitement is naturally high. They are one upping Fractal in the "most powerful processor" game. But that doesn't mean the algorithms are better. If they can pull off instant patch switching with spillover , like Boss, that will put them in a unique category. The thing about their website that concerns me, is the claims of "Artificial Intelligence". Their website and some of their claims seem a bit hyperbolic. There is too little known about it for me to get excited in any way. I expect the feature list to be thin like the other newcomers. The interface looks like the helix. Instead of using touch capacitance like Helix to select effect blocks, they are using rotary stomp switches. That seems too smart for its own good. Should people be stepping on the encoders? Potential for high failure rate. Do you really want to be touching things with your hands that you have also stepped on?

Anyway, like I said, I'm not excited about the QC because it strong suit is profiling, which is a paradigm Kemper has already explored. And there are too many other unknowns. I think a year after release I will have a better picture of whether it is worth $1600.

[Masta' C]

If it does anything, it will cannibalize sales of Kemper. Its basically a Kemper with a nicer interface.

And a quad-core processor that's far and above anything available in current modelers from what I gather.

I agree that tone and feel are "debatable", but the Andertons guys have demo'd just about everything and even they claim this is next level in terms of sound and feel, so that's intriguing at least.

Just to be clear, I'm not arguing. But if Neural can "build a better car", so to speak, I'm all for it.

[Top-L]

Well, that was a pre-production sample and I think they mentioned in passing that it didn't have the complete amp library or fully-featured software that would be available at launch.

Based on the December updates, the current software has 52 guitar amp models (LINK) plus 7 amp captures (LINK)

Basically, there's no reason to sell it short just yet

I'm not selling it short, I just think that once its released, it will sound like the other ones. They can all be made to sound the same with the right IR or EQ. Feel is debatable, some people think the Gt1000 feels better than the Axe FX. Other people say the GT is junk and not at that level. Some people love the way the Helix feels.

Its not going to change the modeling landscape one iota. As its a higher priced device, it wont put any downward pricing pressure on the other ones. And it doesnt have any features the other ones dont. Just the "hope" that it sounds or feels better.

If it does anything, it will cannibalize sales of Kemper. Its basically a Kemper with a nicer interface. (It might steal some customers from Fractal too.)

I'm about to upgrade to a GT1000. You know the feature it has that I really want? It has global blocks, so you can set up an amp block and and EQ block (any block really) to be global across a set of patches. That means, when I tweak the amp and EQ block, I don't have to do it in every patch. This will make patch management a breeze, I can set up one or two amp archetypes that are used by all my patches. If I change guitars, I can just twist the EQ knobs to make it work for that guitar and I won't have to reprogram all my patches. Its the closest to working just like a physical amp and pedals. Because it has instant patch switching and spillover, it doesn't have to introduce concepts like scenes or snapshots.

[Masta' C]

Well, that was a pre-production sample and I think they mentioned in passing that it didn't have the complete amp library or fully-featured software that would be available at launch.

Based on the December updates, the current software has 52 guitar amp models (LINK) plus 7 amp captures (LINK)

Basically, there's no reason to sell it short just yet

[Top-L]

Yes, I've seen that one.

The strange thing is... every demo of it that I've heard sounds like that same bright Marshall, like they've only got a few models in there now.

[Masta' C]

Not sure if it answers any/most of your questions, Top-L, but have you seen this video?

I do know that on his own social media, Rabea has said it's hands-down the best feeling/sounding modeler he's ever put his hands on (which surely includes the Helix, Kemper, etc) and the capture is second to none!

[skud707]

Copying the EQ of an amp or even an effect is base level, entry level processing.

[Top-L]

I believe they have stated that reverb/delay tails will continue between patch switches. And you're right, it's a cutting edge product in some ways, and falls short in a few others - though I think the key is that it falls short in software implementation and not hardware capabilities as that as least heralds the possibility of getting things sorted in future releases. Including things like additional pedals/fx and more amp models, but having profiling and an easy way for community sharing of them goes a long way to "mask" the lack of models it'll ship with - that being said I think it's still planned to ship with ~50 models and several captures on top, so you've still got something like 70 "amps" to choose from out of the box so that's enough to keep me busy for a while anyway.

I think your critiques are fair and I agree that at this point many of the high-end ampsims are down to utterly minute differences and all of them do a very good job. My hope is that Neural once they can focus purely on software dev updates and not mixing time between hardware/software beta testing will deliver updates at a solid pace - they need to PROVE that, for sure, but I am willing to give them the benefit of the doubt for now (hence why I bought in to early adoption).

If there's anything you want to see specifically once mine arrives let me know and I can try it out and respond. But I agree they need to catch up in a few key places. I plan to use mine for studio stuff mostly so the lack of certain "gig" features doesn't impact me much so I'm not as worried about some of those items as you.

Sure, any hands on experience is appreciated. I bet once its released there will be a ton of hands on and unboxing videos.

Its a bit out of my price range. One thing I do like about it is that its a flagship floor processor. The Axe FX3 flagship is rack and needs external controllers. At least this has onboard switches, I think there is a gap in the fractal lineup atm for something like the QC.

My thoughts on profiling are that its not for me, as I don't have many/any rigs I want to profile. But for the right user its a killer feature.


Apparently with the Kemper once you deviate from the base profile by altering gain and eq it gets away from what the real amp sounds like. If they are able to profile the sweep of each EQ knob and gain knob, that would be impressive! Otherwise its just a million snapshots of different amps. If I want a different tone its easier for me to modify EQ/cab/mic/speaker sims than it is to audition thousands of profiles. Thats just me.

I have it narrowed down to the FM3 and GT1000. The Headrush is thin in features and lacks amp switching. The QC is too expensive for me. I don't like how the Helix distortions sound (they are compressed and smooth almost to a fault, but fun to play.) Not many other choices.