This researcher has written some rather interesting papers on modelling the 808 kick, cymbal, and cowbel. They want you to sign up through FB or G+ before you can download from this site, but his work is also readable there and can be found through other channels as well.
I’ve read the bass drum paper. Very interesting read, and a different approach than mine (which is not very “physics informed” - more like collecting lots of data and fitting curves). I more or less walked the same path, but with a machete and 15 MIPS in sight My emulation of “pitch-sigh” (the curve in figure 11), which is the most important ingredient in the 808 kick is done by “simply” taking a phase-shifted, rectified and wave-shaped version of the output and feeding it back into the frequency (and to a lesser extend the resonance) of the state-variable filter that emulates the bridged-T network & its feedback.
The difference in approach might result from the author’s interest in circuit bending. I think his models are explicitly designed to allow the inclusion of the various modifications that have been developed for the original analog circuits over the years.
I do think it’s very interesting that people are publishing scientific papers about this stuff these days. It does make me feel sad about the choices I made in terms of higher education back in the days though.
@pichenettes Your GitHub link appears to be broken.
His approach is the norm in the couple of DAFX/AES papers on analog modelling I’ve read recently (like this). From schematics to (approximative) electronics to DSP blocks…
You mean it’s the approach most favored by academics while your more “brute-force” approach is looked down upon as unelegant and too practical?
@pichenettes is the 808 model you’re referring to the one you used in Peaks? I presume you used a simpler model in the Anushri.
There’s no comparison between Peaks and Anushri.
Peaks uses a model inspired by the 808 circuitry and aims at sounding like a 808.
Anushri’s drum generator is just a synthetic thing - modulated sine wave and noise.
The 808 drum models in Peaks are surprisingly good IMHO.
I wouldn’t say what I’m doing is “brute-force” - it’s just more “data driven” than “model driven”. I’m just not good at the process of plugging in equations, arriving at something for which no closed form solution is available, and then declaring that this or that first or second order approximation would make sense here. Usually I work the other way round - do numerical simulations or record data, look at curves, guess that this or that would be best modelled by a polynomial or gompertz, or a second order model or whatever, and then estimate the parameters without even looking at the physics. This is a process more common in human sciences or in the “big data” world than in physics. It works best in “budgeted” environments when you set yourself constraints: I want to do that with 3 biquads and 2 non-linearity table lookups because that’s my CPU budget. What are the best sets of coefficients / parameters to get there within this self-imposed structure?
This, plus the fact that I’m rarely interested in getting everything 100% right - but more in getting the smallest/least CPU demanding solution that captures the most of the musical “essence” or “usefulness” of the modelled thing. You could call it sour grapes but I like this “blurry edge” you get when you do not overfit.
I hope it’s clear I was joking with the “brute force” label.
Having a background in statistics and research methodology, I personally find your approach the more elegant one. Besides the reason you give above, I also find it more aesthetically pleasing because it somehow feels more native to a computational environment, than trying to “physically model” the behavior of an analogue circuit in software.
Anyway, in the end it’s all about what’s musically useful.
I think pichenettes really hits the spirit here. Do whatever it takes to squeeze the most musical solution out of your hardware. This way great Musical Instruments were made (think of the akward way a ppg or VS works - or even the constraints that lead to the MS-20 - all of them technically horrible but highly musical), if you have to much raw power just boring things happen…… ever played a KRONOS?
Hmmm, I wonder if formal causal inference and causal model discovery techniques for observational data, which are now beginning to be used the human sciences after being pioneered two decades ago by the artificial intelligent community, are applicable to the sort of signal analysis which Olivier describes? Certainly the approach of fitting parsimonious but adequate statistical models in order to gain insights into how things actually work underneath the covers (i.e. insights into the underlying causal mechanisms) is the classic approach in epidemiology, as opposed to automatically-specified, complex, hyper-saturated, typically overfitted and uninterpretable predictive models beloved of data miners. I don’t know enough about signal analysis to answer my own question, but I suspect there is potential cross-over. I started to wonder about this when I encountered the notion of causal functions in DSP - but there “causal” just means correct temporal ordering, I think. Of course correct temporal ordering is also a key requirement for causal inference from observational data - effects can’t precede causes.
@pichenettes yes, I thought the Anushri drums were relatively simple. I think you said somewhere that the Peaks drums were ‘like the Anushri ones, with without the aliasing…’, hence my confusion, probably. I may have misread/misremembered, though.