I'll admit I don't know much about KS but I understand the principle so I made this patch here (kstest.pd), initially with the standard delwrite~/delread4~ construct (and no [block~ 1]). It sounded OK but not great so I set out to learn some more and I stumbled upon this Max/MSP video

which at 9:58 has about the same issue I was experiencing (gritty sound, especially at lower notes). Now, with some tweaking, I managed to make my patch sound nice enough — namely, with a certain amount of [lop~ ]s cutting around each note's fundamental. You can see this in the [pd ks] subpatch — EDIT scratch that, I actually removed the LPs to have a fairer comparison down the line. Carry on

However, the video carries on with using Max's [gen~] which I don't understand very much about but it makes everything sound a lot smoother (see at about 11:50). I tried my hand at something similar with cyclone's delay~ and using a block size of 1 sample in the subpatch but to not much avail — though I'm tempted to believe I'm doing it wrong in [pd samp_ks] using tabread4~ and tabwrite~ but I got told off by PD for creating a DSP loop hooking the output of delay~ to its own input via a damping factor so I figured.

As far as I can tell, there isn't very much difference at the moment between the two patches, at least on the aural side of things: they sound very much the same to my ears! So I was wondering if I'm stuck with copious amounts of filtering or if there's some kind of trickery I'm missing. There is a closed issue in PD's repo regarding this but I couldn't understand much of the theory and practicalities behind the discussion.

Any advice and/or pointers are appreciated

@morpheu5 I've only played with Karplus-Strong once, but I put a low pass filter in the feedback path, defaults to 4500 hz. Not sure if that's where yours was. I also put one on the output, but not so close to the fundamental.

Edit: ha ha, I found where I stole that idea from. He's also doing something with the pluck function.

Now, i've never messed around with 'KS' and don't really follow what is happening, so i might be completely off topic here. But if low frequencies start making things unstable or crunchy in bad ways you might want to try doing what is often done in analog circuits: Shelf or cut low end at the input and boost it back up at the output. The clip sounded very cool, love a crunchy bass-synth

@jameslo a lop in the feedback path is something I tried but I only found it useful to shape the spectrum rather than smooth out the signal to the dac. I ended up with some crazy sets of filters and it's great fun. When I do key-tracked lops I like to put the cutoff either right on the fundamental or maybe at 2x, 3x, depending on how steep the rolloff is. Thanks for the pointer, I'll have a look!

@Booberg Yes, that occurred to me and in effect I'm doing more or less that. Filtering the feedback chain like @jameslo says is also useful in containing unstable signals. However, in this case I'm not dealing with an unstable signal per se. I think the problem really is that the excitation signal has too harsh of an attack and the delay being too long makes the repeats noticeable as almost-separate events. A partial solution at this point is to either smooth the attack of the exciter or low-pass the output signal.

In general, though, I was wondering how using [gen~ ] manages to smooth things out without requiring filtering, and it seems the sample-level processing (as opposed to block-level) has something to do with it.

@morpheu5 Is "sample-level processing" the same as [block~ 1]? If so I don't see how that would necessarily make anything smoother. I wonder if the comparison with [gen~] is really apples to apples. (I also wonder if you think s.elliot.perez's patch is also not smooth)

I was inspired to put something together you might find useful
KStest-Booberg.pd

The big thing there is the impulse generator, it generates one cycle 'waveforms' with a period according to the frequency. I made one that is continuously variable between saw and triangle
In the example the delay time was controlled at control speed, using smooth signals to control the time removes the ugly glitches when changing pitch.
Also, if you want to use noise to exite the loops you might try using brown noise instead to get more low en out of it

Oh yeah, if i would experiment further i might try having a lowpass filter in a loop set at something like 1/8th or 1/16th of the fundamental frequency, and then use a multiplier to compensate for the volume loss (and some transfer function to save your ears when it inevitably blows up!). As you probably know, a lopass filter is basically just a 1 sample delay loop, so you could probably use that fact in some way.. Please share what you end up with

@morpheu5 to me your [pd ks] patch sounds exactly the same as the gen patch if you use the same values he's using: use [1, 0 10( into [vline~], use a [*~ 0.97] instead of [*~ 0.95], and use a [noise~] instead of [pink~]. It sounds brighter but the whole video sounds like the highs are cut to me.
In the original algorithm the [vline~]'s message would be [1, 0 0 $1( where $1 is the period of the frequency in ms. There was also an averaging filter ([rzero -1]) between the read and the write (but then all feedback coefficients have to be multiplied by 0.5)

By the way, this is what I came up with, thank you all for your help! (sorry, the video will actually come up in a few hours)