>He's doing real-time audio processing. So, first, he's got performance issues. In that environment, it can be very helpful to re-use a buffer without having to allocate one (and without a garbage collector). For most functional languages, that's not a place where they shine.

It can be via streaming. The functional portion of the program does all the map and reduce operations on streams. Buffer allocation (and reuse) is automatically handled by the compiler. It's very similar to the way NodeJS hides the event loop. I'm not familiar with audio stuff but I'm sure there's libraries out there that allow you to do extremely efficient audio processing with high level language primitives.

>But, second, he's (probably) got problems like building echo cancellers. The hard problem is, how do you determine whether something is an echo? Writing it in an FP language doesn't help much with that, because the hard part is the definition (and tuning) of the algorithm, not the implementation.

Neural networks can be employed here to "recognize" echos. NNs are basically functional primitives with deterministic inputs and outputs. FP is remarkably appropriate here given that audio is basically a function.