Dither in Mastering: A Deep Dive
The one topic guaranteed to clear the room at any party, explained properly, with science, visuals, and just enough self-awareness to survive it.
Why Dither Exists at All
Digital audio stores amplitude as integers. A 16-bit file has 65,536 possible values for any given sample. A 24-bit file has 16,777,216. That sounds like a lot, and it is, but the moment you do any math on that audio (EQ, compression, level adjustments), you end up with values that fall between the integers. When you round those to the nearest whole number, you introduce quantization error.
The problem isn't the error itself. The problem is that it's correlated with the signal. At low levels, that correlation sounds like distortion, a distinctive, grainy, spitty texture that mastering engineers have described as "like AM radio through a bad coffee maker."
Analogy that actually holds up: If you're measuring rainfall with a ruler that only has centimeter marks, you're going to round everything. That's fine when it's pouring. But at a light drizzle — when the real value keeps hovering between 0.3 and 0.7 cm — you'll keep writing zero, then zero, then one, and the record looks nothing like the actual weather. Dither is like adding a tiny, random wobble to the ruler before each reading. Suddenly your average is accurate, even if each individual reading is still rounded.
Below is an interactive demo. The left pane shows a smooth gradient, think of it as a sine wave fading to silence. The center shows what happens when you truncate it to a low bit depth with no dither. The right shows the same signal with TPDF dither applied.
Notice how the truncated version develops distinct "banding", those are the quantization steps made visible. The dithered version maintains smooth gradation at the cost of visible grain. In audio, that grain is white noise, which is vastly preferable to harmonic distortion.
At 16-bit, the quantization step size is about -98 dBFS. By the time a mix is dithered and quiet passages approach that floor, the dither noise is already below the threshold of audibility, approximately -90 to -120 dBFS depending on the noise shaping used. The distortion it replaces would be far more audible.
The Noise Types: RPDF, TPDF, and Noise-Shaped
Not all dither noise is equal. The shape of the noise distribution matters a lot for how the error is distributed across amplitude values.
The TPDF rule, simplified: RPDF dither is like one die ,1 through 6, all equally likely. TPDF is like two dice summed, you get 7 more often than 2 or 12. That triangular distribution is exactly what you need to fully de-correlate quantization error. This is why "just add some noise" isn't quite the same as proper dither. One die is not enough dice.
Noise Shaping: The Psychoacoustic Trick
Here's the elegant part. Human hearing is not equally sensitive across the spectrum. The Fletcher-Munson curves (or their ISO 226 equivalent) tell us we're most sensitive around 2-4kHz, and significantly less sensitive above 15 kHz. Noise shaping exploits this.
Instead of distributing the dither energy evenly across the spectrum, noise-shaped dither uses a feedback filter to push quantization noise into the high frequencies where the ear is least sensitive. The total noise energy goes up, but the perceived noise floor drops, sometimes dramatically.
TPDF (flat): Noise floor at approximately -93 dBFS A-weighted. Correct. Boring. Reliable. Like a trustworthy station wagon.
Noise-shaped (aggressive): Perceived noise floor can drop to -105 dBFS or lower. Sounds quieter. More noise energy above 16 kHz. Caveat: do not re-apply after conversion.
Noise shaping flavors vary significantly by algorithm: Waves L2's IDR, Pow-r, UV22HR (Apogee), and MBIT+ (iZotope) each have different curves and characters. The differences are subtle in isolation but can compound if noise-shaped dither is applied multiple times. Don't do that.
The classic mistake: Applying noise-shaped dither at every stage of a mastering chain is the digital equivalent of adding reverb to your reverb because you thought the room sounded a little dry. Each pass isn't replacing the previous error, it's adding to it. Apply dither once, at the very last word-length reduction step.
When and How to Apply It in Mastering
Hot take that is actually correct: If you're mastering for streaming at 24-bit and you've been skipping dither because "24-bit has plenty of headroom anyway," you're probably right that it's inaudible. But it also costs you nothing to apply it correctly, and it means the file is mathematically clean before any downstream conversion happens. It's the audio equivalent of wearing your seatbelt on a short trip. Pedantic? Yes. Wrong? No.
Dither once, at the end, to the final bit depth. Everything else is downstream, including this article, and definitely your dinner party conversation.
