Hi, sorry for late response. Thanks for the feedback.
I understand well the true requirement for "transparency" of "analog".
I read this encoding pcm paper. Very interesting is the sample rate and depth ideals proposed.
https://www.meridian-audio.com/ara/coding2.pdf.
20bit rectangular channel (noise free) or lesser depths with relevant noise shaping.
I was playing around with 8bit and different NS curves recently and it is astonishing how well even 8-bit can sound.
It shows me that s/n ratio specs are over-rated because, clearly 8bit already trumps audibility of LP groove noise on its own - that and the fact that most converters capture with a cleaner channel than typical mic in (even the quietest) environment anyway.
So really, s/n can be completely ignored. I'm focusing on accuracy.
The best converters quantize to 24bits, (though technically can only really be 20bit at most internally because of limitations in voltage dividing).
So focusing on depth that is used to mix signals/effects, whether original (quantization) depth is retained, or increased upon entering a new container.
So far people have addressed that accuracy beyond 20 is 'inaudible'.
I'm not trying to prove audibility, merely establish whether increasing depth can increase 'accuracy' of dsp.
If we follow linear 'levels' throughout a 20bit a/d/a chain, addressed using fixed-point (fixed is easier to explain because it allocates levels linearly).
lets say 20bit ladder ADC captured & 'quantized' to 20bit INT.
20 provides +/- 524,287 levels.
This can be sliced further for mixing, for example Pro tools uses 48fixed.
~ +/- 140,737,488,355,328 levels.
mixed down to 20bit(or '24bit') and distributed.
played through 20bit 'ladder' multibit' dac. (i.e PCM1702).
I realize that 48bit is already alot. But to get my point across, using more than 48,
should increase mixing resolution albeit beyond capability of a/d/a but still more realistic approach (I feel).
I'd also like to ask why float allocates quantization in a non linear manner and whether or not this concerns anyone (theoretically)?
As I understand float allocates / quantizes more frequently as it approaches MSB.
If you wanted to be theoretically stubborn, surely fixed with its linear allocation of levels, should theoretically provide a more linear 'signature'.