audio_analysis_test seems to take just over 2s in 2.14.59 non-debug build on laptop (though I think this was without ebur128 due to a mistake when building)

Since when is Leq(m) in it?

I'll check... on main the same test takes 14s...

Removing _leqm->add(interleaved) from main brings it down to 5s.

Using a PeriodTimer around the main pass() loop:

Removed _leqm->add() from main,
main 3.13 2.14.x 1.93 62%

Removed _ebur128->process call() from both,
removed _leqm->add() from main,
main 0.436s 2.14.x 0.378s 86%

Looks like building leqm_nrt with -O2 is about 4 times faster (judging by test.sh in that repo). Probably also worth looking at doing the convolution there with FFT.

Also leqm_nrt seems to loop over channels, so maybe we don't need to interleave the data before passing it in.

See opt branch on leqm_nrt

Merged the -O2 change as e7a04275703b8cd54e1325d42f484ed7d0acab92

Seems to work a lot faster now, at least in 2.17.13 - full length (2hrs) 5.1 audio incl. LUFS/LEQ in about 5 minutes.