So to summarize:
Either with ROON or HQP DSP engines, when using DSD output (>=128) from redbook 44.1 source files, better to use 384kHz wav filters. ACOURATE is OK to produce such files, REW is not for the moment. Right ?
Well I think that this is generally correct.
With Roon, you ideally want both 352.8 and 384 filters if you are upsampling to any DSD, regardless of source PCM rate.
But if you are only using 44.1 then HQP would in theory be better off with just 44.1 filters as it doesn’t have to resample them at all. I believe that in his response, Miska was only suggesting the highest possible filter rate on the assumption that one could send any rate to HQP. Whether or not that is discernible/audible is another matter.
You are right again @nquery, i’ve got confirmation from @Miska:
> If you have source files only in that format and nothing else, then you can create filters for 44.1k sampling rate instead. Normally you would generate filters for highest source file sampling rate you would imagine using.
No problems, but it would still be nice if Roon also fixed it on their end, @brian ?
More specifically, if Roon uses a convolution of say 192 khz, and user has selected up sampling to 384 (or DSD output), then split the up sampling in 2 phases, first to the convolution bitrate, then perform convolution, then (if needed) the rest of the up sampling, then EQ, DSD etc. As far as I can see, this would solve the problem as good as it can be solved when convolution filters have a lower bitrate than the final bitrate.
I’m not convinced there is anything to fix.
There is a downside to adding an extra SRC stage, too, as well as added complexity in SRC planning–both to implement and to explain/defend. You’ve given a really simple example that sounds straightforward, but there are some nasty cases that are not so straightforward–cases where we might convert rates in different directions (44.1k -> 192k for convolution -> 96k for output), cases where we could ping-pong to many rates for multi-stage corrections. It would get pretty yucky, and pretty easy to create a signal path that made no sense at all.
Yes, in general I prefer to avoid filter resampling, but not at the expense of making sample rate conversion something so complex that I can’t explain it to anyone. In the end, filter resampling is a solution to not having the exact matching filter, not a problem that needs to be solved 
Thanks for the explanation, but looks like we will get REW to be fixed instead, just got this email from REW:
Sure, I’ll add 384k as an export sample rate option in the next beta.
Btw, big thanks to John Mulcahy for making and supporting REW, its a pretty awesome piece of software and its free!
2 Likes
@brian: How to setup convolution in ROON when using native DSD source (DSF without any upsampling / conversion) ?
Seems like filter is resampled according to DSD rate with huge impact on server CPU. Not doable with my I5. Any workaround ?
Thank’s.
You can turn off “Enable Native DSD Processing”. This will cause the DSD to be processed as PCM instead, which saves on CPU.
Other than that, there really isn’t a free lunch–convolving DSD at DSD rates is extremely expensive. It needs a big CPU.
[quote=“brian, post:28, topic:30742”]
You can turn off “Enable Native DSD Processing”. This will cause the DSD to be processed as PCM instead, which saves on CPU. Other than that, there really isn’t a free lunch–convolving DSD at DSD rates is extremely expensive. It needs a big CPU.
[/quote] Thank you @brian !
Not a free lunch but DSF (DSD128) convolving OK on my not so big NUC6I5 (6260U) : Processing rate =1.6x
im sorry for digging up this old topic but I was wondering if things changed.
Im currently experimenting with Convolution filters for my Headphones. I Downloaded the 44.1khz and 48khz files from Github, added them to Roon and it works just fine. but I was wondering what the “taps” exactly mean, what are 20k taps exactly doing ? and how is a convolution filter different from a 10 band parametric EQ?
I noticed the Convolution Filters make everything more quite and I need a bit more volume. is it ok to disable volume levelling when using them instead of parametric EQ?
is it better to use the P-EQ for DSD output since I don’t have the matching 384khz files for the C-Filter?
I played around a bit and this is what my signal path looks like in my living room.
my other setup is able to do native DSD512
