Hector Martin

Had a look to the video, and it cheats a bit :). It does increase SNR *only* on a reduced zone of the spectrum (and increases it in other zones). Thus it is effectively reducing overall noise when using e.g. A frequency weighting (to assimilate to human hearing).

But it is not correct (or at least not complete) just to say "it increases SNR beyond 96 dB". It only does it on a fraction of the spectrum (the fraction interesting to human hearing).

That is only half of it. The first demo shows that a tone at < 96dB is still above the noise floor even with flat quantization noise. Shaped dither just improves things further. SNR is already > 96 on a band basis, which the article I linked later goes further into.

I haven't read it yet, but you have to integrate the floor noise throughout the entire band. Calculating SNR against the noise floor is wrong. Will read and come back

Ah, but we aren't talking about SNR, we're talking about dynamic range. Not quite the same thing ;). Music isn't a broadband noise source, so you get some extra perceptual headroom for free.

Let's sum it up, to see if we agree: 16 bit w/out oversampling allow for about 96 dB SNR. Using dithering you can have more than 96 dB SNR by taking advance of human hearing characteristics (an 'A' frequency weighting filter can be used for computations).

Other than that, you can use oversampling to gain more SNR, but that requires increasing Fs beyond nyquist.

You have to be careful though, the lower the Fs is, the more stringent requirements will be put both in the anti-aliasing and the reconstruction filters. Both must be analog and will not be the ideal brickwall filter.