I might have misunderstood what you actually wanted to do, but in principle with a polyphase filter you can directly compute output samples at whatever fractional position you want.I have seen different set ups of polyphase filters where they are not doing the impulse difference calculation I will take a look at that again. I think that is what is confusing me if downsampling and upsampling are the same cost and decimation is not required.
The "gotcha" is that when the output samplerate is lower than the input samplerate, the kernel cutoff needs to be relative to output rate, so the kernel varies by the downsampling ratio in terms of the input rate. If I understood correctly, your solution is to multiply the output samplerate until it's higher than the input samplerate, then downsample by an integer factor afterwards? If that's the case, then sure, you do need a separate downsampling pass in this case (though whether it makes sense to use HIIR rather than another FIR is another thing; the half-band IIRs aren't that great of a fit for modern CPU architectures).
The common theme with any approach though is that things get expensive if you want to pitch up a lot... for which there really isn't any solution other than to either give up quality or to play different samples at different pitches (either by virtue of actual multi-sampling, or by virtue of precomputed mips to synthesize those different samples).
Statistics: Posted by mystran — Thu Jan 23, 2025 11:24 pm