I'm making a fairly complex polysynth that has lots of aliasing so I am upsampling where necessary (complex waves, FM, wavefolding etc) and then filtering.
The butterworth filter from J07.oversampling uses a lot of CPU and I want to keep my polyphony!
4 x lop~ chained one after the other uses a lot less but sounds a bit worse.

What filter are you using for this?

there was a similar thread here some weeks ago.

the ideal antialiasing filter for non integer relations (such as 96->44.1) is more complicated, but for normal /2 downsampling a butterworth topologyis fine.

why dont you go a bit lower, say like 17.6 khZ (nyquist * 0.8)? much room then with 4, 6, or 8 poles.

Yes butterworth sounds good but uses a lot of CPU.
I am usually filtering around 18kHz, I can't hear up there anyway.

Hi Nuromatix try this . I converted the audiolab Butterkreuz filter into automatonism. should be work in all enviromentsbk.pd

Thanks, will check it.

You can get away with filtering at 11025. Sounds way better than aliasing. But you might want to do some exciter stuff above 6-8k for compensation (careful, it's addictive).

The cheap but efficient version is oversample at a factor of 2 and use a cheap FIR to supress anything above 22.05K. Something as simple as this will work just fine if you are already using BLEPs:

This would be the counterpart to [inlet~ lin] :
@bocanegra said:

( could be [block~ 64 1 2] or even [block~ 128 1 2] )
I am wondering why such a linear average is not the common way in down-sampling? Can you elaborate? This vs. Butterworth?

averaging means you only have 3db/oct.

sure, same "issue" exists with poles and zeros for the typical (nonsquared) butterworth.

but i believe the latter is a bit more flexible and does not feel so strange.

try
[slide~ 2 2]
[slide~ 2 2]
[slide~ 2 2]
[slide~ 2 2]

for an 88 to 44 conversion and see if you like the result

@lacuna you are right about the blocksize, my bad... But yeah, my point is that you can get away a lot cheaper than the oversampling example in PD if you dont mind cutting corners. Averaging filters can be stacked for steeper slope and a non linear filter response can be approximated with n taps (cascade of one sample delays with individually tuned gains). If your oscs are polybleps you don't need that much precision in you final anti-aliasing filter.

@Obineg [slide~] is not in vanilla PD. Is it a max object?, and what does it do?

i only have pd extended installed.

for me everything where the user has not added something himself is "vanilla" i am aware that around here that´s used differently.

slide~ interpolates logarithmic between neighbours. the arguments are samples for up and down.