You are quickly becoming the RU-vid expert on profiling and capture technology, thanks for your great work, I hope it results in good things for you and your family.
Leo, amazing again!, this is by far the BEST video explaining the real complexities of making an accurate profile!, keep up the good work, those numbers and the math here are what many misses!
Thank you so much Master Leon. If/ Whenever you want to play something together or you have the time... feel free to send a message: it would be a super honor for me to make a song together!
Great explanation. I'd never thought how complex this technology is. I think we are on the cusp of revolutionary tech breakthroughs for modeling and capturing tech and it will drive the prices down. It's a great time to be alive! Happy Easter my friend 🎸👍
And so the difference between this channel and most others is revealed. Leo Gibson is an actual nerd. We are very lucky to have such a content creator.
Hi Leo. WOW! This vídeo was very nice. And I am very happy to see you really consider opinions of people here (about nonlinearity). The great contribution you did was the training curve. This should be mandatory for all AI builders. IK multimedia, Neural DSP, STL and so one should to see this vídeo right now. I believe in near future there will be a pedal to load these profiles similar to IR's today. I'd like to see anyone trying to transform a profile in a single IR file and try to insert into an IR loader. Great video again! Well done and thanks!
HI, thank you, I'm really happy that you have appreciated the video!!! Yes, actually I really try to read all your comments....and I'm learning a lot from all you...I'm very proud of our community and the suggestions and info we exchange each with the other! Sometimes I'm not able to follow up all your suggestions...but it is just because I miss the time...but, as I said, your comments, support and feedbacks are really super important for me!!
I love how you explain the science of the tube amp and re-amping it..... DA converters to an analog device to an AD converter with limited sample rates..... There is only one relevant test: does it sound good and does it make me want to play (more). The rest is irrelevant.
Hi James, thank you, I'm happy that you like the explanation! Yes, I do agree...it has to sound good and it has to make me wanna play...let me just add, that this is subjective, I mean everybody means something different with "sounds good", etc...therefore these tests try to find the objective things at the basics of a device to sound good...at least in my opinion. Thank you!
Watching this after the NAM phase inversion vid 🙄 I can now only assume it won't be long before something even better comes along. It's great but I'm having commitment issues.
Important video. Thanks for creating that. Another thing that contributes to the null difference from attempt to attempt is that the clocks are not synched. So you are not sampling the waveform at the same time from recording to recording.
Awsome vídeo!! Thank you for great content and analysis. By the chance the raw audio applied in real amp and in the capture was exact the same ? I mean, playing strong or in smooth way would affect the response of dynamic in real amp and simulated amp, so if the raw audio is the same with the same dynamic it we could have pretty close response in both.
Hi, the file used to train the neural network and that goes into the amp is a "standard" file...if I have properly understood the question. Thank you, I'm happy that you like the video!!!
@@LeoGibsonGtr I gottcha! Thank you ! I was taking with a friend mine, if isn't practicable training IA model the all real amp parameters how the Quad Cortex, Kemper and Tonex build a great Amp model? Maybe a mix of tecnologies like wave digital filter(WDF) + Neural Network ?
Awesome video! I am shocked how much closer the ToneX tested than the Kemper and QC in those original tests, at 1/5th the cost. Awesome value for that little device.
Thank you, I'm happy that you like the video! Let me just add the QC and Kemper offer more features compared to the ToneX, they have more effects, more connectivity, etc...that contribute to increase the price of these units.
Profile the gain setting you often use. Then profile the maximum and minimum gain you think you might also use. Then feed these 3 profiles to the app which will then interpolate the rest of the values. I am sure NAM will come out with such feature.
The thing with NAM is that, even though it's about as good as it gets, Steve (the dev) has said it's a very basic and simple implementation of a neural modeller. If NAM is basic, I can't imagine how an advanced version would sound. It would be great to have the possibility to modell the dynamics of a guitar amp chain (or hardware comps for that matter).
Thank you, I'm really happy that you like the video!. I'm not sure about what you mean with "increments of 3": could you please explain better? Thank you!
There is no perfect tube amp so why bother trying to make a perfect digital copy! Take whatever digital box you own, plug a guitar into it and make some music. Leave science to scientists!!!
Great video, thank you for all the work. This only takes into account neural networks technology for modelling. There are other technologies to model systems that are not NN based too.
My amp is perfect for this. It's a Marshall 1974X clone with a tone knob and a gain knob. It's got power-scaling on the on-off knob, but that's a convenience for playing small venues. You'd want to profile it with the power scaling maxed-out. The thing is though, Leo, how many settings do we use on our amps? Surely 90% of us rely on 3 to 5 total combinations of settings? In most gigs, I don't touch my amp's controls throughout the entire show. And I'm sure I'm not alone. Wouldn't most people be able to effectively do everything they already do with a few profiles?
Thank you! I think that the problem is that you would use 3 or 4 settings...but I would use 3 or 4 different setting from yours...and someone else could use 3 or 4 different section compared to mine and yours...that's why I think it's worth to make as much setting profiling as possibile: in order to make a general plugin...otherwise it's gonna be ok for few of us...still without the potentiality of the real tube amp Thank you....interesting topic....
@@LeoGibsonGtr you're right. And I suppose the advantage potentially of a free, open source community is that you could possibly request particular profiles from people. And there's a bell-curve, right? Probably nearly 70% of players would all be happy with the same three settings from popular amps. But I do take your point.
The inability of the neural network to reach the theoretical limit is related to its degrees of freedom. Because the algorithm involves solving sparse matrices (tensor calculations) the degrees of freedom are limited by RAM and processor speed. Its interesting that NAM approaches -40.8 LUFS. I wonder if the Tonex algorithms approach a different asymptote with increasing "epochs". My guess is that Tonex has fewer degrees of freedom than NAM, but who knows. I want a way of converting captures from one platform to another. NAM profiles played on Tonex hardware or a Kemper would be great to try. Another great video, Leo.
Yea, that way we can have an even cheaper device than the Tonex and all the Tones, I just use my PC with my HX Stomp with the Tonex software, but routing it can be a little wonky, might end up getting the Tonex pedal just to simplify that.
Thank you for this high-quality compare! There is many mathematical ways to get NAM model even better ESR, not only calculate model with "standart" architecture with many-many epochs. Here you get flat with many epochs because of learning rate decay. So i think there is ability to get ESR ~0.005 or better. My ears and fingers are not able to guess real amp or nam in blind test if ESR ~0.001. On really high gain amps the best results are higher because of noise (~0.005-0.01) Can you share your output.wav (reamp from v1_1_1.wav) to try to get better ESR?
Thank you so much, I'm happy that you like the video! May I ask you which are the ways to improve the ESR? Any suggestion from you would be highly appreciated! Thank you!
@@LeoGibsonGtr try to use different learning rate decay, by default it's 0.93, try 0.99 or even 0.999. Good practice is to learn model again after learning but with slightly different parameters. For example, you learned to ESR = 0.01, then this model can be trained more but with different "ny" (smaller) and "batch_size"(smaller) parameters. This is not possible in easy mode and from gui trainer
Actually the point of profiling and in general of a Machine Learning approach is to predict the output of a dynamic system withuout having to try every possible different input, doing also surely brings about diminishing returns and can also be counter-productive and prone to training errors. The way forward with respect the AI approach in circuit modeling is always having better algorithms and hopefully new techniques that are less general and more specific to the physical modeling domain. But the "standard" circuit modeling approach can still catch up by maybe leveraging AI in different ways, in the end we are still living a new "Tube Vs, Digital" moment: now we have "AI vs. non-AI Modeling".
Thanks for doing all the hard work, and crunching the numbers. In think it's impractical, and probably unnecessary to have numerous Profiles/Captures. Three basic Profiles - Clean, Crunch, and Dirt, may be all that are actually needed. Algorithms that define/describe the sweep of Control Knobs - Bass, Mids, Treble, Gain, Volume, etc., could be created. I did a little research on them, and they appear to be just simple Potentiometers, with relatively standardized ranges/specs. I realize this is a simplification, and would not produce the absolute "exact sound" of a specific Amp. But, as you showed, an Amp's sound is inherently variable. Variability is something that can be programmed, if needed. And since we don't really need perfect Profiles, they don't exist, we probably couldn't hear the differences even if they could be captured, etc., we could probably make a few specific Profiles that would be close enough to cover the range of the actual Amp, and make it even more flexible by doing it in the Digital Domain.
Unfortunately a "tone control" algorithm wouldn't work - a treble control, for example, doesn't just adjust the amount of treble, it alters the gain structure for the phase inverter and power stage, leading to both amplitude and frequency-derived saturation (distortion). The only accurate way is to sample every tweak of every control.
@@digitalchris6681 I'm thinking that there may be a "sufficiently predictable" line between two known values for any Pot that an algorithm could derive, given sufficient data on how that affect the resulting output. OpenAI, and new versions of GPT may make deriving the Algorithm even more precise. There is currently a way to create an IR from a VST Plugin (perhaps even from NAM) that can then be imported into a Floor Modeler (like HX Stomp) and used as an IR in that unit. Here is a video showing how it can be done: How to make Impulse Responses from guitar cabinet plugins FOR FREE ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-TB5F5pFjXp4.html I don't know if anyone has actually tried this and compared the result in a "Nul Test" to the original NAM Profile/Capture. Would be very interesting to hear the results. Just watched this video regarding cool new features added to the Standalone NAM UI: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-u3mPtCStEP4.html
@@picksalot1 I found the upgraded NAM vid yesterday and will give it a try as soon as I'm able. Haven't tried the home-built portable IRs yet either. One thing is for sure - at the rate software is evolving, by the end of this year the guitar amp & cab emulation/profile/model possibilities will be vastly increased and improved !
Great video Leo!!! I'm an engineer by trade and the quality of your work makes me suspect you might be too. I'm glad to see the comparison of the real amp against itself.
Thank you soooo much, I'm really happy that you have appreciated the video!!!! Actually, I've been graduated in computer science...even if a pretty long time ago...and it is not my main business since many years...but I still love all these stuff... Thank you once more!
Very nice video and very nice explanation! I love to get a bit technical with this topic because it really is very interesting. Thanks for all the info shared! You know what I would really like to see? (maybe in a video or responding a comment) to talk about reamping best practices, how do you set the volume of the signal going to the amp, ground loop issues and things of the sort. These are very important for getting a quality profile. Thank Leo! Best regards!!!
Thank you soooooo much, I'm really happy that you like the video!!!! Actually, I'm working on something much more nerdy...cant' wait to share...even if it's gonna take me some more time...😀
Hi, you are welcome! I would like to make a package of profiles...but NUM is free, so everybody is expecting to get them free....and so it's not gonna be nice vs the ones that have purchased my ToneX Profiles...I'm still thinking...
You are welcome! Let me just say that QC makes a lot of other things, and not only the profiling part, therefore it comes down to what you really need... Thank you!
Cool! Can you do a null test with an overdrive in front of the amp and a profile to see if they react the same? They should if the training input was good.
We live in a wild time. Stable Diffusion allows me to create graphics with my mediocre 2070super that I could never have created by hand. Ozone helps me to improve tracks that I couldn't balance properly in an unsuitable audio space. Sonible plugins and Radix Auto-Align automate tedious tasks, so I can focus more on what I want: the actual mixing process including creative decisions. By the way, ChatGTP translated this comment from German to English, so you don't have to worry about my grammatical weaknesses anymore. Many people are afraid of these technological advances, but I can't wait for the first people with disabilities to make great art through new accessibility.
Yes, I do agree...these are great utilities that allows us to do more and better in less time...my only concern is related to the creativity space...I mean I want creativity to be only a human being area...not A.I. May I ask you what Sonible plugins do you use? Thank you!
@@LeoGibsonGtr You are addressing a philosophical category here. What is creativity? What is originality? Which of these has higher or lower quality under certain conditions? That's not so easy, and I find it very exciting to think about all of that. The Sonible plugins that I use are primarily Smart:EQ, Smart:Comp, and Smart:Limit.
Hi, yes it is philosophical, but super important, as I think our creativity and our art are crucial, and a distinctive factor for us human being...that we have to preserve not allowing AI to substitute us...it's not easy...but it's worth spending time to address it. Thank you for your suggestion bout these tools.
@@LeoGibsonGtr I kinda hope this destroys the current profiler products- it would be so much easier if we could update our hardware ourselves so as to take advantage of whatever software improvements come along: for example the Kemper hardware is - in technological terms - ancient, so by definition the software it runs is compromised.
1) Professional player: Great tube amp,with custom shop guitar 2) Weekend, club player: Tube preamp (Kingsley) , Solid state power amp, (Orange baby) 2x12 cab, Small pedalboard,Player series type guitar. (easiest,safest gig ,most peace of mind.) 3) Home playing,recording: Tube preamp(Kingsley) ,Two notes torpedo cab m plus.,small analog pedalboard. Plug ins, multi effect processors are getting old at monthly rate this days.
I don’t get what the hype about either proteus, or nam plugins is for. For me all the captures i have tried sound horrible. Noone is saying much and noise and hiss are included. How bad the gate works on the nan plug..renders it pretty much useless. Anyone else? I know why they are free..because they are garbage. Just being honest.
Hi, thank you for sharing your opinion. Everybody of course have is own opinion an it's fine...but in the null test NAM just performs better than the other.
@@LeoGibsonGtr I'm gonna stick to tried and tested..mics and amps, vst amp sims, pedal sims. Yes the Nam is better than the proteus, but they are both not good. maybe in future this will change..for now I find both unuseable because i don't play metal and can't hide all that hiss and noise behind a wall of distortion. So many users have done captures..it can't all be user error. I'm gonna try a capture myself of the simplest form of amp I know a rubyamp...just a handful of componenets and a lm386 chip in the netlist, which gets a great sound..that way can rule out user error..if theres hiss and noise then we can say it really is garbage (for now) :)
Totally on point! ... and remember, the guitar and including the cable (or wireless system ) settings also sets the 'Snapshot' flavour ... but the technology is here to stay and expand by everybody ... Developers will be only limited by their imagination ... and inclusive, their imagination would be supported by AI 'Creative Asistants' ... The good, musicians must to Jam a Lot because that is what AI can not replace, the experience of playing together and flow in time with other humans in 'subjective/intersubjective' time ...
Thank you! Let me just say that the input file used to train the network is a standard one, with some strange sounds...I don't think it depends from a specific guitar.
@@LeoGibsonGtr Probably, the guitar model, settings, and player touch wouldn't be relevant variables for the NN grab the amp + recording setup response ... but maybe it would be a nice test for measuring that influence in the chain ... maybe that could bring more nuances about the amp response to the training data set ... I remember the underrated Peavey Revalver ACT (Audio Cloning Technology) tech who took the input signal into consideration, one of the things that the tech try to emulate is to change the input from any guitar to emulate the input from another guitar ... For Example, you could emulate the raw direct line strat 'phase second position quacky input' with a Gibson less paul and so on ... In, practice, it - relatively - worked ... (others developers as Positive Grid included something similar to their emulations with similar results ) ... But It would be interesting to test If the raw input into the NN amp emulation can get the nuances from the instrument with the independence of the data set used to train the Gear ... Because it opens a market for NN emulations of RAW inputs ( something that has a predecessor tech in the Line 6 Variax ) 'One Axe to Rule All of Them' 😂
Have you tried profiling amp+cab with Nam? It didn't work well for me, the profile was very odd, it didn't sound like a cab, but a weird digital amp and very different than the original one
Hi, I have profiled both, I mean with CAB and without CAB...and finally I prefer with CAB...In this video you can here the profile of a Pink Taco v2 in action with CAB: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-fVB_WWsgnHo.html Thanks!
