Exactly, why would anyone want to add dither to an unchanged file? No more bit perfect playback? The purpose of dither (noise) is to randomize the quantization errors resulted from the truncation of a 24 bit file to 16 bit one. A 16 bit file already has dither in it.
A good dither material:
http://downloads.izotope.com/guides/izotope-dithering-with-ozone.pdf
There are a number of reasons why adding dither can be beneficial for a DAC. In practice we are talking about converting 16 bit data to 32 bit data by adding rectangular PDF noise onto the blank (all zeroes) data - its not really dither as such, because we are not talking about treating truncation errors as there is nothing to truncate as its all zero. And the top 16 bits would still be bit-perfect, as one would only randomise the zeroes.
The first reason is small signal non-linearity. By randomising small signals, you improve the performance of the DAC - but this tends to work best when you have a number of DAC's in parallel and feeding each DAC with random noise. By averaging the output, it means you can use much more noise. This used to work well in the past, but stopped making any difference when I invented pulse array, as this DAC technology has no small signal non-linearity. By using noise (I won't call it dither) like this made the SQ smoother and with better depth - all hallmarks of better small signal performance.
The second issue is DSP core PSU noise being signal correlated, which then gets picked up by the analogue electronics, and this mechanism also degrades small signal non-linearity, so upsetting depth reproduction. This issue used to be a major thorn in my side, as the DSP cores I had to manually create from FPGA fabric, and they consumed a lot of power, so there would be lots of noise entering the ground plane. This issue has slowly been eliminated, and since Hugo I have had zero measurable problems from it; this is due to a number of reasons:
1. Dedicated DSP cores on the FPGA, which are extremely power efficient.
2. Dual ground planes making ground bounce no longer an issue.
3. Use of switch mode PSU, with the current return path located under the FPGA, thus meaning no signal correlated currents leaking away from the FPGA section.
4. Nearly a quarter of a second of music data being processed through the WTA, so current draw is less signal correlated.
Now randomising the bottom 16 bits will randomise the DSP currents, and this can reduce this problem as the WTA is busy processing random data, so the signal correlated part gets smaller.
Now in the past I could not hear any benefits in randomising the permanently zero bits, and so dropped the feature. But we are in uncharted territory with Dave, as it offers depth reproduction to a level that I have never experienced before - for the past 20 years I thought 200 dB pulse array noise shapers were good enough for depth - but now with Dave at 350 dB I need to re-examine this issue, as it may now offer some benefit.
Rob