Preserving the original sample points when upsampling, in a non-cheating workflow, would imply having a perfect brickwall filter of infinite length as the interpolation filter. Because otherwise there's simply no way the resampled audio can naturally pass through the original sample points.
IF (and that's a big IF) the megaburrito filter actually does what it claims, short of pulling an infinite computing miracle, it can only do this by bending the interpolated points toward the original sample points in a non-optimal manner.
This can be demonstrated using Audacity and Adobe Audition:
1. Here, a Dirac impulse drawn on silence sampled at 44.1kHz, shown in Audacity which does not perform any interpolation in its display (or, strictly speaking, "linear interpolation"):
As you may know, a Dirac impulse contains all frequencies theoretically encodable by the format, so this is a real test of resampling.
2. Here, the same impulse as interpolated by Audition in display. It has to show the original sample points where they are, so there's your megaburrito filter!
h34r:
See how the interpolated curve passes through all the original points (i.e. all on the middle line except for the impulse in the middle)? Is this really the ideal resampling we're all looking for?
3. I perform offline resampling of the track using the following settings which I believe to be optimal for the present application. Feel free to correct me:
4. The result. Note that the original sample points are NOT on the middle line anymore. (they're close, really quite close, but no cigar.) If keeping the original samples were desirable, why would Audition do it for the real-time visualization but discard it for (supposedly highest quality) offline rendering?
Note also that the pre / post impulse waveforms undulate for further before and after the impulse. You may know this to be a sign of a resampling filter that's closer to ideal, as a mathematically ideal filter would cause the impulse to pre-ring and post-ring for infinite time. (not that this is a bad thing, because proper music samples ought to have all frequencies so close to the transition band filtered out during recording analog-digital conversion.)
To summarize, the fact that
1. Adobe Audition uses a "megaburrito" filter for real time visualization but not high quality offline conversion, and
2. The offline conversion shows characteristics of a higher quality conversion despite not looking like a "megaburrito" filter
Speaks volumes about the actual desirability (or not) of preserving the original samples in integer-multiple sample rate conversion.
Then again, AFAIK no one has actually been able to tap the internal upsampled digital stream of a Schiit DAC for analysis, so who knows if it actually preserves the original samples? :rolleyes: