Quote:
Originally Posted by peeder  BTW rather than padding a bit-shift with 0's I'd much rather they did so with random numbers. Less efficient certainly but we have the cycles. That doesn't accumulate noise, as the signal bits are left untouched, but it avoids distortion. |
Um, random numbers *are* noise. Zeroes are pretty darned quiet. But I think we've beaten this one to death.
Dithering works because it adds (carefully-tailored) noise *below* the LSB of the sample, and that statistically pushes some of that information up into the length-reduced sample.
What I think you're missing is that while the random numbers aren't adding noise to the particular signal, they certainly *will* add noise to the overall mix, because those bits are significant on other channels, which makes it unlike dither.
I think the real issue is that if you take a quiet signal and boost the heck out of it, it'll sound increasingly like crap regardless.