There hasn't been much discussion around Convolution lately.
Fairly new to Convolution filters and going though the process of using Convolution to create linear phase EQ. Running Room EQ Wizard to take Freq Resp measurements, Holm Impulse for phase (at 1m distance from each speaker) and Importing/Plugging in the EQs from REW into RePhase to generate the Impulse WAV files. Is working quite well - post EQ measurement is close to predicted target in REW when using the Auto generate PEQ function (using Constant Q, linear-phase in RePhase for actual filter creation). Also using the Filters Linearization/Paragraphic Phase EQ in RePhase to flatten the phase response (at 1m distance) - quite an improvement in sound quality over basic PEQ.
The issue I have, is that 8k filter taps work to 20khz, but if I generate 65k taps, there is a sharp rolloff at 10khz (with exact same filters). Why does increasing the taps (which is supposed to be more accurate) roll-off at 10khz?
I've used RePhase quite a bit, and I've occasionally measured strange results above 10KHz, but (for me) they've always been the result of an error in measurement technique.
Are you seeing that roll off in a measurement in REW or Holm, or in the modelled response in RePhase?
I ask because sometimes when the time zero isn't exactly correct on a measurement, the measurements can show fictional frequency roll and phase wrap at higher frequencies (because they're more time sensitive). The longer filter is more accurate, but adds much, much more latency to the measurement (65K taps is close to a full second of added latency in 44.1), making it more likely that the time zero autodetection might be off. If your roll off is based on measurement, check on the time zero of your impulse, and see if you still have the same problem once you've lined up the time zero. In Holm the "largest peak" setting is usually the most reliable time zero auto-finder.
Other things it might be:
1)If you're using a USB mic, the mic has a different clock than your output device and high frequency info (especially high frequency phase measurements) can be a little less reliable, especially in high latency systems (in my experience).
2) Sampling rate: you didn't mention what sampling rate you made the filters at. Are you measuring at the same sample rate you made the filters for?
If you're not seeing the problem in a measurement, but rather when you make the filter in RePhase, that's a stranger case. If that's the case, if you'd be willing to post a screenshot of RePhase when you generate the filter I can try and help troubleshoot.
Generally speaking, you can feed the PEQ filters REW generates into JRiver's PEQ module, and then just use a RePhase impulse filter for phase correction on top of that. It produces identical results to doing it all in RePhase (in my experience), but with a much, much lower latency convolution filter (phase adjustment alone generally requires fewer cycles than frequency + phase adjustment).