[Storyville]

To get dithering, you have to get bits, and by extention, samples. Think of a recording as a film camera taking very rapid photographs. In film, you have a number of frames per second - and they move by so fast you can't actually tell that you are seeing a rapid succession of still photos. Your brain just says "hey, it's a moving image." Same thing with sound. Sample Rate is in essence the number of sonic images that occur over any given amount of time (usually per second). Therefore, a Sample Rate of 44.1kHz, means you are hearing over 44 thousand different blips of noise in one second. Bit Depth on the other hand is the sonic equivalent of resolution. This determines the quality of each sample. Generally speaking, this has a much greater effect on the overall sound quality than the sample rate.

Computers function on discrete mathmatics. This means that nothing in digital truly exists on a continous scale. Everything is based on increments, including the amplitude of any given sample. Let's say we zoom in on one sample (one 44 thousandth of a second if we go by the standard sample rate). That sample either has an amplitude of 1, 2, 3, 4, 5, 6, 7, etc. There is no amplitude 3.14159265. Bit Depth determines the number of numerical steps of amplitude you get in any given sample, and the overall largest value you can reach.

16 bits may look something like: 1, 4, 7, 10, 13, 16
24 bits may look something like: 1, 2, 3, 4, ...., 19, 20.


Ok, so REDUCING BIT- DEPTH

Audio CDs play at 16bits. If you are going there and you recorded in 24-bits, at some point you gotta get to 16. Bit reduction is easy. Your computer just reduces the value of each sample by the necessary percentage (I don't know what that percentage is). This is called quantization. Let's say we were going to a bit depth that required a 50% loss of numeric value. If you have a series of samples that goes: 0, 3, 6, 10, that series gets converted to: 0, 1.5, 3, 5.
But one problem: the value 1.5 does not exist. We are only working with whole numbers here. Therefore, the value of that sample experiences quantization error. Depending on how the converter functions, that value is going to either be truncated (rounded down) to "1" or rounded (rounded to the nearest integer) to "2." Either way, we do not have exactly the correct value.

In and of itself, this one instance of quantization error would be inaudible. It would just fly by too fast. However, sound functions cyclically and the way we hear is based on repeating patterns. In practice, bit reduction yields a cyclical quantization error that functions in relation to the audio content. This results in space monkeys. Weird glitchiness is heard because the exact same quantization error is occuring in the same places.

Luckily, there is a process called dithering. Dithering randomizes how quantization occurs - whether it is rounded or truncated. This randomization breaks the cyclical nature of the quantization. So rather than getting space monkeys, we essentially get white noise.

All of this occurs most prominently at the edges of black holes. Meaning at low levels of volume. Think about it. If your series of numbers is 5, 4, 3, 2, 1, and you reduce everything by 50%, you get: 2.5, 2, 1.5, 1, and .5. Truncated yields: 2, 2, 1, 1, 0. Rounding yields: 3, 2, 2, 1, 1. The quantization error, relative to the original value is a very high ratio. Five seperate values is now equal to only 3. 2/5 of your sound is now quantization error!

So dither when you fade in or out, and dither when you downgrade bit depth.