The term dither is used to describe a special form of wide-band noise that is added to a digital audio signal after the wordlength is reduced. The purpose is to randomize quantization error of low level signals.
In digital audio, the original analog signal amplitude must be encoded in discrete "steps." The number of possible steps is determined by the wordlength. For high level (large amplitude) signals, the "step size" relative to the signal is small and the resulting quantization error caused by "rounding up or rounding down" to the closest discrete step is small.
At very low signal levels (small amplitude); the error becomes quite significant and the effect is a clearly audible distortion that increases in level as signal level decreases to the least significant bit (lsb) level. By randomizing the error; the distortion is eliminated and this results in a perceived increase in the dynamic range even though the addition of dither (noise) actually decreases the measurable dynamic range.
There are a number of different types of dither; with "Flat" and "Triangular" or "HPDF" (High-pass Probability Density Function) used commonly in digital audio. Both distribute the noise energy fairly evenly across the audible spectrum. Flat dither is preferable for encoding audio which will undergo further processing and Triangular dither is generally preferred for final encoding and wordlength reduction. An additional process known as noise shaping can be applied to effectively "shift" the noise energy away from the midrange where the human ear is most sensitive to the bass and upper treble regions.
For additional information; also see: