Curious about comparison to the LADSPA plugin

[jacksongoode]

Thanks for implementing this! I was curious if you tested to see what the performance is like compared with the repo's LADSPA plugin?

[vaisest]

On my 5800X3D Audacity Windows system a 7.5 minute audio clip takes 36 seconds with this non-LL model and the LADSPA plugin takes closer to 1.5 minutes. I don't really know if this is a fair test. I also remember the LL and regular model having much different performance, and I'm not sure what the prebuilt LADSPA .dll uses.

I don't see why this should differ in performance, though. Practically all of the processing time taken is the model processing

RNNoise recompiled with AVX takes around 12 seconds for the same file.

[jacksongoode]

That's interesting... I think the prebuilt uses the LL model without any lookahead. I'm not sure exactly why the model is larger (the paper's don't have a lot of insight on this) but the evaluation might involve seeing how much of the ring buffer gets filled per period in the LL vs non-LL model?

I see you used 480 over as the period size, so the latency should be identical I suppose (40ms?)