Dither And Noise Shaping

Canonical Definition

Dither is intentionally added noise used before a quantization or word-length reduction so quantization error becomes statistically decorrelated from the signal. Noise shaping uses feedback or filtering to redistribute quantization-noise energy across frequency. Dither linearizes low-level quantization behavior; it does not increase headroom, restore clipped samples, or improve the analog hardware.

Why It Exists

Without dither, low-level quantization error can repeat with the waveform and appear as harmonics or signal-dependent distortion. A suitable dither distribution makes that error noise-like instead. TPDF dither is a common choice because it can remove the signal dependence of simple quantization error under its intended conditions.

Dither Versus Noise Shaping

Process	Main job	What it does not do
Dither	decorrelates quantization error	remove noise or recover clipped detail
Noise shaping	redistributes quantization-noise spectrum	eliminate total quantization noise
Noise reduction	reduces an existing noise component	serve as a synonym for dither

Correct Placement

high-precision processing
→ final level and filtering
→ dither / noise shaping
→ intentional word-length reduction
→ no later gain or processing that invalidates the dither level

If another truncation occurs later, that later reduction needs its own analysis. Adding dither after every DSP block only accumulates noise.

Measurement

use low-level tones and inspect harmonics plus absolute noise
state input and output word lengths
state dither distribution, amplitude, correlation, and shaping
confirm the analyzer and analog path are quieter than the effect being measured
check out-of-band shaped energy against converter and amplifier bandwidth