I don't have time to read all these posts... but I saw one thing I wanted to clarify a bit.
You'll most likely see an improvement from working at 24 bit dynamic resolution -- but I recommend
against working at 96 kHz --
unless you're outputting to a 96 kHz format.
The reason (without getting into details) is that the conversion from one bit depth to another does not degrade the sound any more than it would have been if you had started with the lower rez. And if you're performing any manipulations in the higher bit depth (EQ, mixing, compression, other dynamics) at all, you'll probably see an improvement in the final product over performing the same manips at lower bit depth.
(Every added bit of storage capacity added to a stored number increases the possible stored values by double. A 1 bit number can store 2 values, a 2 bit number can store 4 values, a 3 bit number can store 8 values, and so on, to where an 8 bit number can store 256 values, a 16 bit number can store around 65,000, and a 24 bit number can store around 17
million values.)
Now, by contrast, doubling your sample rate gives you double the values (say from 44.1 to 88.2) which, if you then perform a bunch of manipulations might help
some before converting back -- but a non-even multiple causes the whole wave to be
remapped ... if you've ever resized bitmap graphics you've probably noticed that resizing by exactly double (or half when you're dealing with even numbers of pixels) produces a cleaner picture than resizing by, say, 201% or 49%. Sound works the same way.
Working at 96 kHz and then doing an uneven resample down to 44.1 kHz can actually leave you with
worse sound than if you stayed at 44.1 kHz to perform a given set of manipulations.
Not only that, adding bit depth gives an exponential increase in dynamic resolution (a 20 bit number can store 16 times the values of a 16 bit number -- but only adds 25% to the overall storage requirement. They used to call that stuff "mathmagics"...
) -- while doubling (or more) the sampling rate doubles (or more) the storage requirement for a given sound.
Now, 44.1 kHz can provide (as the Nyquist theorem states) usable frequency resolution up to about 20 kHz... but if the dynamic resolution is less than adquate you'll still end up with hashy, harsh sound. (If you've ever heard 8 bit sound, you know what I mean.) Since high frequency soujnds in normal sound usually take up a tiny part of the dynamic resolution [high freqs are the tiny little bumps that ride the big [more bassy] waves in a mix, for instance, and give detail and sheen] adding resolution -- even without increasing sample rate -- can provide a much more
accurate representation of those high freqs.