-
Regarding the comparisons with other modelling amps/pedals, I wanted to offer some perspective. I just finished a PhD in AI and Music (2020-2025) where I got some experience in how these modellers are trained, so I might be able to offer some insights (or not).
Early approaches to modelling relied on impulse responses (IR) which look at how a "click" (impulse) sounds after being sent through a system. This could be an amp, speaker, microphone, room or whatever. If you can play and record a click through it, you can get an impulse response within reason and then build a system to mimic or replicate that response. The problem is that it's hard to isolate these individual elements to get a "perfect" IR - for example, if you try to capture the IR from an old tube amp how do you make sure you don't introduce additional colouring via the microphone or the room which both impact the result. This is compounded when it comes to playback because you take the output of the IR processing and pass it through another speaker with its own IR that colours the sound again etc.
The result is that older modellers required a lot of careful calibration followed by a high quality full-range, flat response speaker to get the best results, but even then it sometimes doesn't "feel" like a real amp.
The tone master amps do two things - one is that they no longer rely on IRs, the other is that they include the physical cabinet and speakers in the modelling process. First off, the newer convolutional neural networks (CNNs) can do a similar job to IRs but with a much wider range of possibilities. Instead of modelling an amp as a linear system (which its not), the CNNs can reproduce highly dynamic, *non-linear* responses which is part of the big step up in realism. The other smart move from Fender (IMO) is that by providing a good pine cabinet and known speaker they've isolated those variables and can refine the sound even further. I've been really impressed with the tone master amps I've played so far and I think the standardization of the cab and speaker choice is a huge factor in their sound. Remember, if you're trying to model the performance of a known system (i.e. an old tube amp) then it makes sense to keep the number of variables to a minimum by keeping the cabinet and speaker to be (roughly) equivalent.
That setup allows Fender to train their CNNs (one per amp model) as a "black box" where guitar audio as input produces the same output as an old tube amp would. This is still a developing area of research, especially around how it responds at different settings, but the general idea is that you have thousands of hours of (clean input, amp output) pairs to train on and then let it run until the output is sufficiently "close" to the real amp for the same input audio.
The other method (which I don't think they use yet) is "white box" modelling where some or all of the key electrical components are modelled by smaller CNNs. This approach would get you closer to the "mix and match tubes, swap out transformers" ideal but it would require an enormous amount of training data to model those components successfully. There are some other approaches (e.g. Wave Digital Filters WDFs) that try this but the idea is the same. It's hard to get right but I think people are getting closer every year.
Understanding these principles might help to understand the tone master range and the strengths and weaknesses. For example, knowing that they did extensive testing to model the sound as a complete package, I wouldn't rush to swap out the stock speaker. Also, the training data was most likely clean guitar sounds so it doesn't make a lot of sense to use these amps as a pedal platform if you use a lot of modulation on your input signal. On the other hand, if you know that you like the straight sounds from one of the original amps then the tone master should get you pretty close in a lightweight and more consistently priced(!) package.
Regarding repairability, the main component is indeed an ICEpower 100w class D amp board and I'd expect equivalent replacements to be available for some years to come. The other parts are slightly more complex than regular solid state amps (e.g. ARM cortex processors instead of shark DSP chips) but we don't seem to have the same reliability concerns about other studio gear like keyboards or interfaces. I get the romantic/pragmatic angle that a tube amp is techinally repairable anywhere in the world (as long as there are tubes) but it doesn't mean that the tone master is bad per se.
-
07-10-2025 07:02 AM
-
I’ve used one of these live a few times now and been impressed, more than sufficient volume wise for straight ahead jazz. The DI functionality makes it very handy indeed for musicians playing a wide variety of gigs.
Sent from my iPhone using Tapatalk
-
Many thanks for the detailed information.
Originally Posted by xavriley
-
That’s interesting - are there any sources describing how they do it? I’m curious because lookup tables are linear systems and tubes are famously non-linear in their response Originally Posted by InsufferableRhythm;[URL="tel:1416643"
-
Old thread, but... Originally Posted by BigDaddyLoveHandles
I recently did a pit gig with another guitarist. Young kid, kinda green to the scene, but a real nice guy and a good player (surprised me how good). He had the Guitar 2 book. Well he came in with a TM Twin. Took up WAAAY too much space behind us in the pit, we kept trying to rearrange everything to hear ourselves right and fit that amp into the tight squeeze. Finally I brought my spare Quilter for him to use, and it worked out great.
-
Originally Posted by InsufferableRhythm
Here’s yet another classic example of the kind of information dispute we should all want to see resolved. Which is correct? Does Fender use CNNs or not? If so, is this what drives Tonemasters? Is this info even available outside of Fender’s proprietary documents? What is the source of each of the opposing opinions? Is there support that makes one right and one wrong? Originally Posted by xavriley
-
My scientific mind agrees with you and wants to know how stuff works. My guitarist’s mind doesn’t care as long as stuff inspires me to play. Originally Posted by nevershouldhavesoldit
-
That ain't support, my friend. Support is data, documentation, and demonstration. Data are king, but strong support from reputable sources ("expert opinion") may be the best available evidence where data are lacking. True experts are not self designated - they earn the title through peer review of their work products and processes. They have the documented knowledge, skill, and experience to have earned trust in what they say. Originally Posted by InsufferableRhythm
I'm not saying you're wrong, since I don't know the correct answer to this question. But you're putting your entirely unsupported statement ("I am correct") out there as though your Nobel prize speaks for itself. At least the author of the opposing point of view has a PhD in a relevant field (assuming he's telling us the truth, which is not guaranteed although I have no reason to doubt him). But you're going to have to do better than a righteously indignant "I am correct". How about some evidence?
-
Originally Posted by InsufferableRhythm
I'm too old to be...
-
I'm not sure that these two views are so far apart.
One poster said he didn't think that Fender used white box, but he seemed to allow for the possibility. He seemed to be assuming that the components inside the white box would be modeled with individual CNN's.
The other poster said that they do use a white box approach, but, if I understand it, he said they don't use CNN's inside the white box, but, rather, traditional DSP techniques to model individual components.
Use of table lookup to specify a non-linear function makes sense to me.
The comment about avoiding speaker swapping makes sense to me, but tbh, and hijacking the topic for a moment, I've never quite understood how to pick one that you're going to like, for any kind of amp. I can't evaluate a piece of gear without trying it in multiple settings that I play in regularly. And, that's not so easy with swapping speakers.
-
Credibility must be earned. Anyone with a phone or an internet connection can post anything on any forum. And it happens every day. Just saying "I'm correct!" provides zero credibility, and failing to show any evidence tends to destroy any that may have existed. Maybe you did some work, but no one else has any idea what work, or how it was done.
-
Digital is better than tubes in a lot of pragmatic areas. And I'm not sure how many of us could pick out which amp is the original tube amp and which is the TM behind a curtain. Realize there is some disagreement in this area. In any case, what you're buying with these is hardware and software that not only sound close but, and quite importantly, provide functional operation as close to the original as possible. Knobs work the same. Tone stack. Loudness. Distortion vs. clean. It's not only supposed to sound the same, but work the same. Fender is selling into their users experience base that goes back 75 years and innovating would be contrary to that marketing approach.
-
The person playing it can tell the difference if they know what they are doing. Most people don't know what they are doing or they wouldn't be buying a piece of crap digital amp.
Why no one can admit these are disposable throwaways is beyond me. It's the fan boy flavor of the month digital gear today that'll be completely obsolete tomorrow. It's like asking who has a good pizza and someone saying "dominos". Digital acid-reflux.
Fender is selling overpriced junk as usual but there is definitely some genius in it because a sucker is still born every minute who will buy one judging by how many of these threads I see. Digital doo-doo that is cheaper to make and with a higher profit margin. $1529 for the tone master Twin, 1129 for the princeton. How to get less amp, for more. Yup makes sense and iT sEz fEnDuR oNnIt.
-
Hello IR.. interesting stuff. I'm finding that the more recent low cost modelers from China (e.g. Joyo, Mooer) are better to my very subjective ears than earlier modelers from credible sources like Strymon. They just sound better. Not as cold with more warmth.. Anyway, I sort of understand basic Impulse Response tech concepts (though I could not do the convoluted math) but have no idea how this tech has recently evolved and where it is going next. And I'm probably going to remain a tourist looking from afar and simply enjoying the benefits. But as I said, it's interesting stuff and with the mix of technical and non-technical participants we have here, there are going to be an array of responses when these discussions ensue. Originally Posted by InsufferableRhythm
As for being asked to show your work, you are, of course, free to respond as you like but I wouldn't take it personally or even as an attack. It's not meant in the academic sense you may be used to. It's more the tradition of 'oh yea? If it's 6V then show me the link where you got that..' Forums have always been like that and 'go find it yourself' is a perfectly OK response. At least here we tend to be civil. After all, some here are great guitar players with decades of jazz behind them. Some are pretty good engineers (broad application of the term) with insights into applying gear to making music. There are jazz guitar historians. Some are crazy pedantic and can talk music theory until we get them cookies and a nap. And some perform jazz in public for real every week and remind us of what is actually important. There are even a few who have quite a bit of hands on experience with amp modeling and like the tech. However, not many can look under an amp modelers hood and it would be enjoyable to have more insights into that.
-
But, the "Fender Tone Master Twin" is a great sounding amp, the two speakers add more resonance than a single speaker, IMHO.
-
Stop being a musician. Originally Posted by GuyBoden
But, the "Fender Tone Master Twin" is a great sounding amp, the two speakers add more resonance than a single speaker, IMHO.
Sent from my iPhone using Tapatalk
-
Excellent idea. Originally Posted by Christian Miller
Edit: You're correct and there are less expensive hobbies, for example Chess. I'll only need to buy a cheap Chess set or play online free.
Last edited by GuyBoden; 07-14-2025 at 05:56 AM.
-
Apologies - I didn't want to cause a fight
I was giving my perspective as someone who is (relatively) on top of the trends in academic literature regarding amp modelling. I was making a bit of leap regarding CNNs because the tone master came out around the same time that CNNs became viable for amp modelling, and I thought that the timelines made sense.
I'm prompted to re-evaluate my comments following IR's response though - to be clear, it's entirely possible that Fender are using white box modelling (i.e. emulating individual components/groups of components) and just doing a really good job. Initially I doubted that because the academic literature hasn't had a lot of success in that regard, but maybe there's some unpublished secret sauce that allowed them to make a breakthrough. I did have a look on LinkedIn and there are definitely some Fender employees that have the EE/DSP history that would make this plausible. There are many routes to the top of the mountain, and the only thing that really matters is the end result.
That said, I'm still confused about how IR came to the conclusion that there are lookup tables involved - you can't (in general) tell much about the implementation of a complex function based on outputs alone. It would be trivial for a CNN with enough parameters learn the same function as a lookup table and you wouldn't be able to tell the difference. If we also consider that the tone master uses a rather large quad core processor, this seems incongruous as it would be a huge overkill for lookup tables. That tilts me toward machine learning models being involved because there is no way that Fender, in all their business wisdom, would overspec the most expensive component in the amp! Again though, this is just my semi-informed opinion. I would be delighted for a Fender engineer to come in and explain how wrong I am. I'd be equally delighted to hear more about IR's reverse engineering work but I also respect that it's an unreasonable ask to expect them to do the work of detailed explanations for my sake. In any case, I'm glad that there are other people trying to understand this stuff.
CNNs or not, I just wanted to reiterate my point about tone masters modelling the amp as a complete physical unit including the cab and speaker. I think this gets overlooked in the other discussions that I've seen.
Also, engineering these things requires a testing cycle. You make a prototype, take some measurements and then make some tweaks. Rinse and repeat. The way CNNs are trained is very similar in fact - you start with a prototype (a random collection of numbers, representing thousands of small audio filters), take some measurements (run the audio through those filters and compare the spectrogram with a desired example), then make some tweaks (adjust the numbers in the filters to get closer to the target i.e. gradient descent). The only difference with the CNNs is that this process is fully automatic and can be run millions of times a day. This gets a bit more complicated if you have to physically reproduce the sound and record it, but it's still possible. It could be that the real breakthroughs here are in the testing cycle and the analysis tools that let us compare spectrograms with ever increasing levels of accuracy.
-
Thank you for your friendly, well thought out, and clearly expressed explanation. You offered more than basic support for your opinion and based it on sound logic and information understandable to all. Most appreciated is your willingness to be proven wrong. Originally Posted by xavriley
Apologies - I didn't want to cause a fight I was giving my perspective as someone who is (relatively) on top of the trends in academic literature regarding amp modelling. I was making a bit of leap regarding CNNs because the tone master came out around the same time that CNNs became viable for amp modelling, and I thought that the timelines made sense…
That said, we still don’t know with certainty what’s in Fender’s box
-
yeah, just like guitar Originally Posted by GuyBoden
Excellent idea.
Edit: You're correct and there are less expensive hobbies, for example Chess. I'll only need to buy a cheap Chess set or play online free.
otoh there are gazillion of chess books (and unlike guitarbooks many are actually useful), there are nice chess sets, online courses, premium memberships, chessbase software, more books etc.
just like guitar
-
I don't know if Fender used black box or white box technology developing the Tone Masters, but if memory serves mine came in a brown box.
-
Originally Posted by xavriley
I think there might be some confusion between convolution as used in DSP and convolutional neural networks (CNNs). They’re not the same thing.
Fender’s Tone Master amps use convolution in the traditional signal processing sense. That means things like FIR filters for cabinet sims and reverb, not machine learning models. CNNs are used in AI, and they’re way too heavy on processing and latency for real-time audio like this. The measured latency on Tone Master is around 2.4 milliseconds, which wouldn’t be possible if they were running neural networks.
Also, Fender hasn’t said anything about using AI or neural networks in any of their materials. If you’ve got a source on that, I’d love to see it. But everything out there points to traditional DSP, just done really well.
From a hardware standpoint, there’s no GPU, neural processor, or high-performance compute core inside these amps. They use embedded ARM chips with limited processing power and memory, which are perfect for real-time DSP but not capable of running neural networks efficiently.
Appreciate your post, just wanted to clarify that part.
-
Appreciate the clarification. I didn’t realize TONEX was running a small LSTM on a Cortex-M7. That’s good to know. It shows that some simple neural networks can run on embedded chips if they’re small enough and well optimized.
That said, I still think the original post might be mixing a few things up. A 16-unit LSTM probably isn’t processing raw audio waveforms. That would take a lot more processing power and a much bigger model. It’s more likely that the LSTM is working on a simplified version of the sound, like tracking dynamics or tone changes over time. The actual audio is probably still handled by regular DSP.
Fender’s Tone Master amps look like a different approach. Everything they’ve published points to traditional DSP, not machine learning. Just wanted to help clarify that.
-
Thanks for the response. I think part of the confusion is the phrase “the LSTM is the source of nonlinearity.” That makes it sound like the model is replacing something like a waveshaper or saturation stage and processing all the audio on its own. That’s a big claim, especially on Cortex-M hardware.
A 16-unit LSTM is very limited in what it can represent. At 44.1kHz, it would struggle to model detailed signal behavior across time and frequency. What seems more likely is that the LSTM is modulating a few key DSP parameters. Things like gain, filtering, or dynamic response. The actual audio shaping would still be handled by DSP.
Saying the audio is “processed through it and directly used” skips over all of that. Unless there’s a teardown or source showing the model is acting directly on the signal at the sample level, this sounds like an oversimplification. Happy to be educated here, but everything sounds vague and pretty optimistic.
-
Because that is just your hobby horse and as of yet there is no evidence to back you up. You may end up being right; you may end up being wrong and the Tonemaster amps might be very reliable and run for decades. There are millions of digital electronics that work just fine for decades, even more. I have a several that have not reached the end of their life cycle after 30 years, and we will see how they do. on the other hand, I made toast at lunch today in a toaster given to my parents as a wedding present in 1956 and which still works great. Originally Posted by DawgBone
The most reliable amp I have ever owned thus far has been a Roland Cube 60. Thousands of hours, hundreds of rehearsals, 100+ gigs, nary a hiccup; my circa 1970 Fender Pro Reverb, however, has had to be recapped twice after failing on gigs- and I probably only ever used it on about 50 gigs. Based on my experience, the modern stuff is far more reliable than the classic point to point hand-built tube amp. The enemy of electronics is heating and cooling cycles, which happened abundantly in tube amplifier technology. Solid state and digital technology, on the other hand, generates far less waste heat and does not stress components and solder joints near as much.
I am reminded of similar arguments I have heard from car enthusiasts who think anything built after 1959 is a cheap piece of junk. Carburetors and points that you can set with a dime are the automotive equivalent of tubes. However, if you got 100,000 miles out of a car built in 1959, you were doing really good; most of them were in the junkyard long before that. 100,000 miles on a car built this year and it will just be broken in. The world does progress.
Reply With Quote
“Shearing style”
Today, 05:26 PM in Comping, Chords & Chord Progressions