Karplus-Strong with Jaffe-Smith sample rate-independent

Question

I tried to implement the Karplus-Strong algorithm with some of the Jaffe-Smith extensions (improve pitch, dynamics, etc). Here is the code.

When I run it with a sampleRate of 48000Hz, I get very reasonably sounding string, without any hissing, aliasing or very high frequencies.

The thing that is messing me up is that when I increase the sampleRate (to 2x, 4x), I get a hissy/high freq sound. This is a 48k example and a 192k example.

I'm pretty sure I've messed something up on my implementation, as I was learning about this while reading the papers and writing the code. So please feel free to point anything else that I may be overlooking.

If I simplify the algorithm to be only a single lowpass filter (a filter like to the one I use for the random input) with the N-delay (which, as far as I understand, is the original algorithm), it works much better (without any difference between sample rates). As soon as I add any of the modifications (like the "pitch fix" on module C), it messes up.

My current guess is that increasing the sample rate increases the available frequencies. But I don't get how they are not being filtered like all other frequencies by any of the low pass filters.

The brown noise input is not part of the original algorithm, but it was one of my many attempts to figure out what's causing those extra frequencies (I was assuming that a higher sample would imply a higher frequency noise, which would make the sound naturally higher pitched). It didn't help much, but I left it there because I thought it makes sense.

Any comments, suggestions or ideas are appreciated.

update

The modification on the Karplus-Strong are the ones described in this paper. I do a simplified version, that shows on my top level comment. The relevant parts:

Instead of using a delay of N = round(sampleRate / freq) we use a modified N and include the C block ($H_c$ in the paper) to compensate for the rounding and improve how close the pitch is to the desired frequency. This is the first part of the update function and the cn line on the step.
I also apply what the paper call "Dynamics" ($H_d$). As far as I understand it's another lowpass filter that tries to balance the energy on each partial. On the code, it's the second block of _update that calculates R and the last step of the step function.

If you remove those two, you get a trivial Karplus-Strong with a delay line and a lowpass filter. When I add any of them, I get a similar sound to the examples above.

I'm pretty familiar with Karplus-Strong, but could you tell us essentially what the Jaffe-Smith modification is to the Karplus-Strong algorithm? I looked at your code, but I couldn't tell from comments where to focus on the Jaffe-Smith extension to it. — robert bristow-johnson, Jun 18 '21 at 01:55
okay, so they put on a first-order APF for fractional delay to tune the loop to exactly the period length you want for the specific note. i know about that. the APF has different delay for different frequencies so the higher harmonic partials won't have the same loop delay that the fundamental has. so they can get a little detuned, which might actually maybe sound "better" in some cases than if they were strictly harmonic. — robert bristow-johnson, Jun 18 '21 at 05:33
now, at higher sample rates, the integer-sample delay is more samples for the same note. and the variation in delay that you get with a recursive APF will be less and the harmonic partials will be less detuned from their harmonic frequency value. — robert bristow-johnson, Jun 18 '21 at 05:36
if you want all of your harmonics to be strictly harmonic, i might suggest a linear phase FIR fractional sample filter. then the behavior should be about the same for different sample rates (except you should get more higher harmonics using a higher sample rate). — robert bristow-johnson, Jun 18 '21 at 05:38
Here is an answer spelling out how to do the linear phase fractional sample delay the delays each harmonic partial by the same amount — robert bristow-johnson, Jun 18 '21 at 14:16

hotpaw2 · Answer 1 · 2021-06-22T00:33:47.143

1

An increased sample rate requires proportionally longer filters (both IIR, FIR, and FFT), in terms of the number of samples, for the same filter transition widths, flatness, and stop band attenuation. IIR or recursive filters can also become more sensitive to numerical noise (quantization, rounding, etc.) at higher ratios of sample rate to cut-off frequencies.

I suggest increasing the length of all your filter kernels and implementation sizes proportionately, and if you are using a short data type, try longer ones.

edited Jun 22 '21 at 00:33

answered Jun 17 '21 at 03:53

hotpaw2

35,346
9
47
90

does this mean that whatever is on those papers (and, well, ALL papers) is always conditional to a certain sample rate range? Is there a reason why this is not explicitly stated? (like: this approach makes sense up to X Hz) – fserb Jun 17 '21 at 13:12
1

I think the papers might assume (unstated) that one will try to pick a size of filter (etc.) appropriate to the sample rate. So there is no simple limit to the sample rate given enough computing horsepower (GFLOPs) and memory. – hotpaw2 Jun 17 '21 at 16:12
are we talking about the feedback filter? what filter? – robert bristow-johnson Jun 18 '21 at 01:57

Karplus-Strong with Jaffe-Smith sample rate-independent

1 Answers1