A little later: haj did write at least one article about PDL and sound-processing - I'm listening as I write this to the very nice bit of music he has at the end.

For audio input, I'd note that a physical limitation of doing DFT on a finite, short window of discrete input is that it creates artifacts from cut-off. This is mitigated in various ways, and the easiest way in PDL-land is to use PDL::DSP::Windows. See jjatria's Advent article for more.

More generally for real-time-ish stuff, PDL doesn't yet have a very fully-tested real-time capability. I still intend to experiment more fully, but one approach might be to set up a "flowing" transformation (so you pay the setup cost only once), then keep updating the input sample then reading the processed output. If anyone does have a go at that, I'd love to hear your findings!