Comment by unlinked_dll | Hacker News Reader

Comment by unlinked_dll | Hacker News Reader

Parent Full thread

Also a guitarist. I own 3 amps, two tube and one hybrid. Used to have a solid state. You really can't tell the difference? It's night and day to me, even with the newer Fender mustangs that have great models.

I won't turn in my HT40 for a solid state until Kempers can deliver the same quality at the same price (which I picked up off craigslist for $400). Now my little mono price 15W (which is loud as hell) sounds pretty terrible in most settings but I haven't found a solid state that can give the same character with the right settings. Plus real spring reverb, i can't kick a DSP to make it jangle.

If you're talking acoustic sure solid state wins every day of the week. But in the mid tier amp market for electrics tubes win out.

I do think though that most people don't know how to use a tube amp to get a good tone though. Crank the power amp volume and use your gain as a volume control, push the input with a clean boost or light drive to get distortion. The harmonics stack through the chain, and that's how you get the crystal clear sound without being drowned in noise or fuzz. It's also something you can't do on a solid state, since the input is buffered and the DSPs don't really handle that idea well.

I'm just an upright bassist.

I firmly believe that DSP could emulate any guitar amp. If it's real, then it can be measured, and if it can be measured, then it can be simulated.

What I think is the big deal is, an amp might be capable of producing a certain tone quality, but that tone quality is unreachable if the controls don't cooperate or are not intuitive. It's a user interface problem. What's more, if it's frustrating to the guitarist, they will notice even if the audience (or the bassist) doesn't.

I also think it's a non-problem to some extent because tubes are not prohibitively expensive, and guitar amps don't need to be exceptionally powerful.

> I firmly believe that DSP could emulate any guitar amp. If it's real, then it can be measured, and if it can be measured, then it can be simulated.

This is absolutely true, but the possibility of it doesn't tell you anything about someone having been successful at it.

One way to think about is that the market for the product can be split into three populations: Those who will buy even if the simulation is poor (they can't tell, don't care, or buy based on price or other wizbang features), those who won't buy no matter how good the simulation is (they don't care about the cost/weight/power, are happy with the classic stuff, ones snub simulations for social reasons, ones who prefer products without long boot times or with physical knobs, or just can't tell and are going to be conservative), and those who would buy but only if it's good.

I think in many cases this last group is much smaller than the other two. As a result it doesn't necessarily make commercial sense to make the simulation especially good.

Even in the case of linear effects (e.g. typically reverbs) where it's easy to measure them and determine scientifically that any difference would be (nearly-)inaudible there are many implementations that don't even bother to try being particularly faithful. For non-linear effects like guitar amplifiers it's much more difficult to objectively characterize their performance.

[I own no tube amps or other similar things, but I understand why people do.]

Indeed, and an additional issue is that success is impossible to measure. Objective testing (double blind or whatever) is prohibitive, and there's a tradition of judging gear based on subjective opinion, that will be hard to overcome. We could already be past the stage where the simulations are good enough, and never know it.

Commercial success might not be driving the development of these products. Anybody who's that good with DSP could make more money outside of the musical instrument business. Someone is doing this because it's their passion.

There are plenty of folks who research this problem at the manufacturers as well as in academia. Stanford, McGill, NYU, UMiami, Aalto University, are a few places where you can find great researchers. The leading journals are the jAES and DAFX journal, and in industry I know of Fender, Native Instruments, Waves, IZotope, among others that were hiring PhDs for this kind of role in the last year. Decent money, not SV money however, but it's more honest work than going to study audio projects for DARPA contractors.

> those who would buy but only if it's good. > > I think in many cases this last group is much smaller than the other two.

Maybe, but the professionals have spoken and deemed the modeling amps "good enough". Lots of them now regularly tour with a modeling amp so they don't have to carry a half-ton of equipment and risk their precious vintage amp in the hands of stage and transport monkeys.

And, I might concede that tube amps sound better than solid state ... if you have one of the "stage queens". Marshall's and Mesa Boogie's, for example, are notorious for having a few exemplars in a line that sound great, but the vast majority of them are just kind of "meh".

The modeling amps allow you to always have the "stage queen".

It's a deceptively tricky DSP problem. The surface issue is that nonlinearities create harmonics that wrap around the Nyquist frequency, which creates aliasing. To deal with it you oversample, which introduces some frequency/phase distortion, and in the cutting edge solutions they run at 8-16x oversampling (which means 8-16x more CPU cycles for the distortion algorithm). And it gets worse with heavier distortion.

On the algorithm side there has been some crazy good work with creating black box models using Volterra Series and dynamic convolution, but it's an O(n^2) algorithm (although I believe there's an O(logn) solution out there). The top of the line modeling amps (mostly Kemper, there are a few others) use these approaches among others to give you hundreds of amps at your fingertips, and they sound great. Plenty of high profile bands use them, Metallica probably the most prolific.

I bet someone has applied convolutional neural nets to the problem, since I'd suspect they're exceptionally well suited for black box modeling of nonlinear distortion. Sub in an atan(kx)/atan(k) for your activation function and each neuron would be it's own first order model of a tube stage. The asymmetry might be more difficult to handle, which is where a lot of the warm tube goodness comes from.

The UI problem is well known. That's really what sets one amp design apart from another. No tube amp is particularly unique in signal path or component choice, the ability to dial in a tone is the secret sauce behind the classic designs imo, also why some newer models from Marshall are so much worse than newer guys on the block, but that's just my opinion.

> which introduces some frequency/phase distortion

I agree with everything you said here, except this. Can you explain this? Why does oversampling introduce distortion?

Because upsampling requires interpolation. I'm using distortion in the abstract/academic sense here.

Any signal through a non ideal (or nontrivial, eg a gain) system will introduce distortion, harmonic being new energy at different frequencies, frequency distortion in the relative magnitude at the output that is different than the input, and the same for phase.

If you use a polynomial interpolator there will be harmonic distortion. If you use an interpolation filter there will be frequency distortion, which is highly correlated to the phase distortion.

What that distortion is and if it is perceptible depends on your design and constraints. But the key element that is at play here is that oversampling for nonlinear system modeling, which introduces harmonic distortion after the fact, is that it can amplify frequency (and therefore phase) distortion and your constraints are tighter than they normally are.

If you want an example take a the ideal interpolator, which is a sinc function. Because a true sync is non causal you cannot implement it, therefore lossless interpolation is not achievable. There will always be some loss of information, the design problem is trading off that loss with the resources required. We can do pretty damn good upsampling today however.

One thing that might work in our favor here, is that for this particular application -- simulating a guitar amplifier -- we are only dealing with the subset of behaviors that are physically realizable in an electronic circuit.

Reply on news.ycombinator.com