Hi all,

I'm working at a new project involving pluck~
(one of the Karplus-Strong algorithm implementation in PD)
Something I don't understand (see the attached patch) is the
fact that every time that pluck~ is triggered with a new note
the sound is sometimes "full" "warm", other times "thick", "small".
(let the metro bang for 10-20 times...). I'd like the sound being "full"
every time the note is triggered. There must be some sort of
"phasing" inside I'm not able to find (Maybe the delay lines?)..

Any suggestion is welcome...

Cheers,

AlbertoZ

http://www.pdpatchrepo.info/hurleur/help_with_pluck.zip

This problem is inherent to the synthesis method. There's a couple of things that need explaining...

1. Because you use noise, and a very short envelope slice of it, there is no way to determine exactly what spectrum the excitation pulse will contain. It could sometimes have a lot of high energy, or sometimes a lot of low, depending on a random influence.

2. The delay buffer decays exponentially, which means it never truly goes to zero once excited. Residual low frequencies and noise persist in the buffer for a long time.

Combined, these two factors make the implementation very unpredictable. Both the initial condition of the buffer, and the precise spectrum of the noise pulse have an effect on the evolution of the sound. Sometimes the noise components close to the resonant frequency will be out of phase with those already in the buffer. Sometimes, by chance alone, the appropriate components will be very low in the first place.

Two things need doing to correct this problem.

1. The buffer should be cleared before a new note (which will take at least one period to zero fill the buffer in Pd) This can be achieved with some kind of "note off" command, and a [sig~ 0] connected for a short time.

2. A more precisely tuned noise source must be used, one that is certain to contain at least a minimum energy in the proper band (close to the period of the delay buffer). This can be acieved by using more than one seeded noise source and a pre-filter to ensure it is distributed in a Gaussian fashion around the resonance point.

The way to implement an improvement on the KS algorithm, so that you have no delay introduced by zeroing, and a consistent tone, is to pre-fill the buffer with properly distributed noise. AFAIK, there is no easy way to do this in Pd and writing a [pluck~] external would be best.

Thx Obi for your quick and precise answer..

I'll try to follow your suggestions:

1. Maybe an idea for clearing the buffer could be done "just before" the note on command.
   Yes, it takes one audio block period but hopefully this is negligible to the hear.

2. Regarding the nouse excitation source: yes, a check on the minumum energy of noise
   that enters the delay line must be done. I'll investigate the buffer prefilling,
   a set of precalculated noise arrays with the proper energy could be a good idea.

I'll try with the following: a "note on" message triggers the following events:

1. buffer clearing
2. load the KS delay line using a table containing the properly-shaped noise source
3. start the recirculation in the delay line.

This should take 2 block periods. Hopefully this can be done without going to externals...

One of the few VSTi that I bought few years ago was HarpTime (http://nusofting.liqihsynth.com/harptime.html).
I made several trials and the output is extremely repeatable.
I know that it is physical modelling (but in this case KS IS physical modelling??),
but I'd like to have something at least the basic of it in PD.

Cheers,

AlbertoZ

Yes, KS is "physical modelling". The term is useful to distinguish from spectral modeling. With spectral modelling you try to emulate a sound by directly following the spectrum at some points and with physical modelling the implementation tries to copy the physical process going on, to some extent, following the propagation of forces and displacements, boundary reflections and phase inversions..etc. But there isn't one way of doing PM.

KS is an example of a waveguide model, as opposed to a discrete MSD (mass-spring-damper) system. It is really a kind of IIR (recursive) filter.

I have seen two interesting string pluck implementations in Pd. The proper way, (as done by P Cook and JO Smith) uses two delays, one for each direction of propagation, and two filters which represent the end reflections.

Syntax the Nerd made a nice pluck unit using the fast t3_envelope and small blocksize in the Bot collection.

If the duration of the excitation is greater than or equal to the buffer propagation time, and a copy is passed straight through to the output, then the noisy attack leaves no perceivable delay at the start of each note. Unless the blocksize = 1 it gets hard to tune high notes.