Hi there,
I'm pretty new to pd and I'm trying to figure out how to make a convolution reverb patch that can be used with incoming audio from [adc~] (ie the impulse response would be applied to the incoming audio).

Ben Saylor's excellent [partconv~] external is perfect for the job. Unfortunately, the patch needs to run in an RjDj module and since RjDj uses Pd vanilla I can't use any externals.

Does anyone know of any patches or abstractions or other info that could be helpful? Would this be do-able within a pd patch? Or did Ben Saylor create the [partconv~] external because it isn't?

Thanks!

Here is an article describing partitioned convolution:

http://pcfarina.eng.unipr.it/Public/Papers/188-Mohonk2003.pdf

The article gives an impression of the complexity. In Pd you have some handicaps. For example there is no object to do zero-padding, and workarounds would make the process inefficient, if possible at all.

Thank you very much for the information. Would it be possible to do it without using partitioning? For my purposes, I'm open to greater latency than usual. Thanks!

I think this brings up a good secondary question - at what point is it actually easier to hard code something in C than to make some monstrous PD patch?

I love PD, but I've seen some people fail a bit making octopuses that would live better in the command line realm.

@sthulin said:

Would it be possible to do it without using partitioning?

Never tried it, but it might be possible. For a start, I've attached a patch showing how a signal block can be zero-padded and Hann-windowed.

To process data in frequency domain without changing the phases in Pd, you would use something like [block~ 1024 4] for a 1024 point FFT and 4 times overlap. But with zero padding you have to take care of the overlap yourself, since you use only half of a signal block each time. You could store the signal in a delay line with [delwrite~] and [delread~].

Also note that Pd does not have a centered FFT, time zero is supposed to be at index zero so if your filter kernel is centered it will have alternating phases in frequency domain.

If you make progress, please post results.

Katja

Edit: forgot one thing, if you do non-partioned convolution with long FFT's you may get audio dropouts because there is no way to spread the workload. So before you start working on this, better first check how long FFT's you can do on your target platform.

http://www.pdpatchrepo.info/hurleur/zeropad.pd

Thank you very much for your responses - this is extremely helpful. Admittedly, much of this is over my head at the moment as I don't have a background in digital signal processing, but I'm keen to learn what I can and will hopefully be able to work with someone who knows a bit more about this than I do. I'm trying to come up with something for the end of February, and will post if I make progress.
Sam