You can't simultaneously say that you're dithering to represent low amplitudes while also saying you're keeping enough samples to capture all audible frequencies. Dithering doesn't create resolution out of nowhere, it sacrifices temporal resolution for amplitude resolution. It's also bad for compression (hence why modern video encodes are done at 10-bit even for output to 8-bit devices), and worse if you want to use a source as the basis for further work (i.e. a remix). If you want to store your signal in a simple, convenient way, and not have to carefully tweak the levels for each individual recording, 16 bits isn't quite enough. And as the article admits, extra resolution certainly can't hurt; worst case is the extra bits are thrown away.
Also 44.1KHz is a pain to do in realtime (there's not enough headroom to really filter out the higher frequencies without damaging the 20KHz response), meaning you need a separate mastering step which is inconvenient and frankly unnecessary. 48KHz is a much more sensible standard to work with.
192KHz may be dumb, but 48KHz/24-bit is perfectly sensible. It gives you fewer ways to make mistakes than CD-quality (44.1KHz/16-bit), and at some point the extra space is worth that, particularly since it may well compress better than a dithered 16-bit signal.
> Dithering doesn't create resolution out of nowhere, it sacrifices temporal resolution for amplitude resolution.
Could you elaborate on that? As I understand it, dithering trades distortion for uncorrelated noise, under the assumption that the distortion is more objectionable than the noise. Where did you come to the conclusion that temporal resolution is affected? The temporal resolution is instead related to the frequency resolution.
I agree that 48kHz/24-bit is the most sensible for production, especially since (as you said) you don't have to worry about the levels too much. But when you master a track, you pay very close attention to the levels anyway, so those 8 extra bits don't do you any good. I think most people can't hear past 12 or 14 bits anyway, unless the audio has a particularly wide dynamic range.
The article talks about representing signals below the notional noise floor using dithering, which requires either temporal dithering or something morally equivalent to it - if your ear is detecting an average of dither + waveform, then it has to have several samples to average from.
Also 44.1KHz is a pain to do in realtime (there's not enough headroom to really filter out the higher frequencies without damaging the 20KHz response), meaning you need a separate mastering step which is inconvenient and frankly unnecessary. 48KHz is a much more sensible standard to work with.
192KHz may be dumb, but 48KHz/24-bit is perfectly sensible. It gives you fewer ways to make mistakes than CD-quality (44.1KHz/16-bit), and at some point the extra space is worth that, particularly since it may well compress better than a dithered 16-bit signal.