Good explanations here.
I recall having similar concerns about dithering when I started Rendering wave files in WL 3.0 many moons ago !!!
However, whenever a digital audio file is “processed” and saved as a newer version of an older file, dithering is generally a very good idea, as it actually works to reduce distortions in digital audio material. Most of the impact is with the “least significant bits” (LSBs) in the processed (e.g. mathematically manipulated) digital audio. These are generally the lowest volume data points sampled in the files (which may well be the quiet parts or musical “rests” within the music). The purpose is to help remove the Analog to Digital “confusion” that happens at these low signal levels, where the conversion to an LSB of “0” or “1” may be difficult for the processing to determine and settle upon. This indeterminate LSB “state” can and does cause audible distortions during playback, which is what led people to incorporate “dithering” to digitized audio signals.
A very interesting side effect to adding dithering, is that it has been found through human A-B testing to extend the dynamic range of audibility below the actual level of the applied dither and bit limits of the 44KHz-16bit media, improving the audible noise floor of the recorded signal! It appears to partly be a psychoacoustic phenomenon that human hearing provides (penetrating through noise to selectively hear other sounds within the background), but its been statistically proven to work in a number of human research tests run over the past few decades.
In other words, if I recall correctly, a 44/16 CD’s theoretical dynamic range is 96 dB [20 log (2^16)]. By carefully adding shaped dither, this dynamic range is actually enhanced beyond this theoretical dynamic range limit, as people can actually hear dynamic material below the 96 dB point on digitally recorded material! When you realize that the human hearing apparatus is capable of about 130 dB dynamic range, 96 dB doesn’t sound like much. However, in most recordings, it’s very rare for the noise floor to fall below 65 dB. Even so, the sublties of dither do make a difference. Hence, whenever a file is manipulatively processed, dither is necessary to reduce distortion and improve audibility.
The real question becomes whether 24 bit files really need dithering as the dynamic range of 144 dB [20 log (2^24)]actually exceeds human hearing limits? The answer is “yes”, because it still has the capacity to reduce audible distortions caused by the indeterminate LSB issue discussed above.
So, at the risk of repetition, whenever an audio file is manipulatively processed, dither is necessary to reduce distortion and improve audibility.
FWIW, Arjan: There is at least a dumb question, the one that’s not asked!!!