Every mastering engineer deliberately adds noise to your music. Not accidentally. On purpose, as the last step before export. If that sounds wrong, you haven't met quantization distortion yet.
What Happens When You Reduce Bit Depth
Digital audio represents amplitude values as integers. A 24-bit system has 16,777,216 possible values per sample. A 16-bit CD has 65,536. When you reduce from 24-bit to 16-bit, every sample gets rounded to the nearest available integer in the smaller system.
That rounding error is called quantization distortion, and it has a property that makes it sonically worse than ordinary noise: it is correlated with your signal. The distortion is not random. It is a deterministic function of what the signal is doing at each moment. At high amplitudes this is inaudible, because the signal-to-error ratio is high enough to bury it. At low amplitudes, where a slow fade approaches the noise floor, you stop hearing a fade and start hearing artifacts: a grainy, harsh smearing that sounds nothing like an analog fade-to-silence.
The theoretical dynamic range of a 16-bit system is roughly 96 dB (the precise figure from the quantization SNR formula is 6.02 x 16 + 1.76 = 98 dB for a full-scale sine wave, but 96 is the common shorthand). Twenty-four-bit gives roughly 144 dB. The headroom is not the problem. What fails at the bottom of that headroom is the character of the error.
Dithering: Trading Distortion for Noise
Dithering adds a small amount of random noise to the signal before quantization. The word "random" is the entire point. By decorrelating the quantization error from the signal, you replace deterministic distortion products with something that has no relationship to the audio content.
Random noise has a perceptual property that distortion lacks: it is separable from the signal in human hearing. A quiet signal buried in random noise remains a quiet signal plus a noise floor, still perceivable if faint. A quiet signal buried in its own distortion products just sounds harsh and wrong. Stanley Lipshitz and John Vanderkooy at the University of Waterloo formalized much of this in their 1987 paper "Dither in Digital Audio," published in the Journal of the Audio Engineering Society, establishing the mathematical basis that recording engineers had been intuiting for a decade.
The counterintuitive result: dithering allows signals to be perceived below the theoretical noise floor. A fade that reaches -100 dB in a 16-bit undithered file sounds like silence or digital artifacts. With dither applied, the signal modulates the random noise floor, and human hearing integrates that noise over time, extracting the underlying signal from it. The fade genuinely trails into perceived silence rather than falling off a quantization cliff.
TPDF vs RPDF: Not All Noise Is Equal
Two probability distributions are commonly used for dither noise.
Rectangular PDF (RPDF) generates uniformly distributed random values, with every amplitude in the range equally probable. RPDF dithering eliminates harmonic distortion but introduces noise modulation: the noise floor changes character depending on signal level, audible as a subtle pumping effect tied to the signal envelope.
Triangular PDF (TPDF) generates noise by summing two independent RPDF sources. The probability distribution is triangular: central values occur more frequently, extreme values less so. This extra step completely eliminates noise modulation. The noise floor becomes constant and signal-independent regardless of what the audio is doing, which is exactly what mastering requires. TPDF dithering adds about 4.77 dB of noise compared to undithered audio, but that noise is stable, random, and perceptually far more benign than correlated distortion products.
TPDF is the baseline that nearly all modern dithering algorithms start from or replace with something more sophisticated.
Noise Shaping: Moving the Noise Where It Matters Less
Flat dithering spreads added noise evenly across the frequency spectrum. Noise shaping takes that noise and redistributes it to match the psychoacoustic sensitivity curve of human hearing, pushing energy into frequency ranges where our ears are least sensitive.
The mechanism is an error feedback loop. The quantization error at each sample is fed through a filter and subtracted from the next input sample. This pre-emphasizes the input in a way that cancels the shaped noise in sensitive frequency bands while amplifying it above 10 to 15 kHz, where the auditory threshold rises steeply. The net effect is a reduction in perceived noise even though total noise power has not decreased. The noise is simply moved to a region where hearing is comparatively insensitive.
Lipshitz, Vanderkooy, and Robert Wannamaker published "Minimally Audible Noise Shaping" in the Journal of the Audio Engineering Society in 1991, providing the theoretical framework for psychoacoustically optimized shaping curves. The practical consequence: noise-shaped dithering can offer perceived dynamic range well beyond what raw bit-depth arithmetic predicts.
What Commercial Implementations Do
POW-r (Psychoacoustically Optimized Wordlength Reduction) was first made available in 1999 by a consortium that included engineers from Lake Technology, Weiss Engineering, Millennia Media, and Z-Systems. It ships in three tiers:
- POW-r 1: TPDF dithering with no noise shaping
- POW-r 2: noise shaping across a broad frequency range, adding roughly 5 to 10 dB of perceived dynamic range
- POW-r 3: aggressive shaping optimized for the 2 to 4 kHz band, where human hearing is most sensitive, adding up to 20 dB of perceived dynamic range in that critical region
UV22HR (Apogee Electronics) takes a distinct approach. The original UV22 algorithm appeared in 1993; the High Resolution version followed in 1999. Rather than conventional psychoacoustic noise shaping, UV22HR concentrates dither energy at approximately 22 kHz by placing a narrow high-frequency bias in the digital bit stream. This keeps most of the added noise above the audible range and well above the ear's sensitivity peak, without following a standard psychoacoustic filter model. The result is spectrally quite different from POW-r: relatively flat from 1 to 18 kHz, then sharply amplified right at the top of the band.
When and How to Apply Dither
Apply dither only when reducing bit depth, specifically from 24-bit or 32-bit float to 16-bit. If you are exporting at 24-bit for archival or further processing, dither is unnecessary. At 24-bit, quantization noise sits around 144 dB below full scale, well below the threshold of any playback system.
Dither must be the last process in the chain, after all EQ, compression, and limiting. If you run dithered audio through further processing, you re-quantize it and the dither noise gets quantized along with the signal, defeating the purpose. Apply it once, at the end, when bit depth is being reduced.
Never stack dither. If your DAW exports with dithering enabled in its settings, disable the dithering in your mastering plugin, and vice versa. Two layers of dither add noise without adding benefit.
Which algorithm? POW-r 2 is a reasonable default for most material. It handles a wide range of content without the pronounced high-frequency noise accumulation of POW-r 3, which can become faintly perceptible on headphones at loud playback levels with certain program material. UV22HR is a legitimate choice for acoustic and classical recordings where that high-frequency characteristic sits comfortably within the program. On highly compressed and limited pop masters, the differences between algorithms become negligibly small.
The differences between dithering algorithms on well-mastered material are subtle. The difference between properly dithered and undithered 16-bit audio is not subtle at all, particularly on material with dynamic fades, reverb tails, or low-level ambience. That is where quantization distortion shows up clearly, and that is why the noise is added on purpose.