Super useful for machine listening to stick this into an FPGA (and why
not start baking standardised versions into SoC) because so much of
the code for these projects is in this tedious pre-processing that is
common to almost every application. It's also the place I see the most
errors - and you don't want to find you trained a whole set of shit
with avoidable mistakes in your input DSP, because that invalidates
everything.
