The mathematically ideal reconstruction filter is known (sinc) but impossible -- or at least inconvenient -- to realize. Therefore, approximations are used in DACs, with a careful weighting of what's more relevant vs. what is less relevant in practice.
When investigated thoroughly it turns out that the frequency response is the dominant factor. NOS and "Slow" filters have significant roll-off of the treble (up to several dB's by the time they reach 20kHz) and that is clearly audible as long as your hearing is still good to, say 15kHz and above... with source material that has enough content there (white noise preferred here as test signal).
The attenuation at fs/2 is way less important, as is the general attenuation above fs/2. Time-domain aspect (post-ringing only vs. pre-and post-ringing, or in other words, the phase response) is also not that important. And the higher the sampling rate the more so:
- Music seldom contains significant amounts of energy around, let alone above 20kHz.
- Any imaged components (frequencies mirrored at fs/2) are above normal hearing range for most people. With 48kHz or greater everything is 100% out of the hearing range of anybody in the world.
- Any downstream components that are not total junk will not produce any relevant additional intermodulation distortion products from the mixing of the images with the original content. And if they are junk, the normal IMD from signals below fs/2 is already dominating the picture, a bit more junk added doesn't have any real impact anymore.
- Any normal music signals also do not contain single dirac samples which would actually excite the ringing. But even when it is excited, it is at fs/2, so above normal hearing range.
Go
here for a demonstration what different filters would sound like if we could hear beyond 80kHz.
Bottom line. Don't worry too much about DAC filters, and use the "Sharp" types which have the flattest frequency response. But even when using NOS, you can apply a bit of EQ to make it flat up to 20kHz and then the ususally reported differences don't expose themselves anymore.
For ADC's (Analog to Digital Converters), the filters are way more important as insufficient filters acctually corrupt the signals unrecoverably, adding aliased signals (stuff above fs/2 folded down below fs/2).
