Dither converts correlated quantization distortion into uncorrelated white noise, making it perceptually benign

Without dither, quantization error is correlated to the signal: as the signal changes, the error follows a pattern producing harmonic distortion audible even at very low levels. Adding a small noise signal (dither) before quantization randomises the rounding decision so the error becomes white noise — uncorrelated with the signal. The perceptual consequence is significant: distortion and low-level ‘granulation’ noise are inaudible at typical listening levels, while white noise at -93 dBFS (16-bit TPDF dither level) is below most listening environments’ noise floor. Dither trades a signal-correlated artefact for an uncorrelated one at the same average power — a favourable perceptual exchange. The technique also means dithered audio can encode sub-LSB information in a statistical sense, resolving the claimed ‘time resolution’ limitation.

Examples

TPDF dither = sum of two uniform random distributions ≈ white noise. At 16-bit: dither adds ~-93 dBFS noise; undithered quantization distortion at -80 dBFS signals is audible ‘grunge.’ Practical use: dither when reducing bit depth (24→16) at mastering; never dither the noise-free special-case sine test signals.

Assessment

You are reducing a 24-bit mix to 16-bit. Compare the audible result of (a) plain truncation and (b) TPDF dither, for a signal playing at -60 dBFS. Which sounds better and why?