What is dither? When to use it in mastering, and when to skip it
Dither is a tiny, steady noise added when you reduce bit depth, so the quietest sounds fade into noise instead of distorting. Here is what it does, measured, and when you need it.
Dither is a very small amount of random noise added to audio at the moment its bit depth is reduced, for example when a 24-bit or 32-bit float master becomes a 16-bit WAV. Without it, the rounding that the conversion needs leaves an error that follows the music, and on quiet material that error is heard as grainy distortion. With dither, the same error becomes a steady, featureless hiss far below the music.
What problem does dither solve?
A digital file stores each sample as one of a fixed set of levels. A 16-bit file has 65,536 of them, so its smallest step sits about 90 dB below full scale. Your DAW works at far finer resolution, so when it writes a 16-bit file every sample has to be moved to the nearest available level. That move is the quantization error.
On loud, busy material the error is small and random enough to behave like noise. On quiet material it does not: a reverb tail, a fade-out or a soft piano note only crosses a handful of levels, and the error repeats with the waveform. Repeating error is distortion, harmonics that were never in the recording. Dither adds a little random noise before the rounding, which breaks that link: the error stops following the signal and turns into a constant noise floor.
What it does to a quiet sound, measured
We quantized a 1 kHz sine tone at -80 dBFS, roughly the level of the last seconds of a fade and only about three 16-bit steps tall, to 16 bits twice: once by plain rounding, once with standard TPDF dither. Levels are per FFT bin, 4 seconds at 48 kHz.
| Result at 16 bits | Plain rounding | TPDF dither |
|---|---|---|
| The tone itself | -80.1 dBFS | -80.0 dBFS |
| 3rd harmonic (3 kHz) | -111.9 dBFS | Below the noise |
| 5th harmonic (5 kHz) | -107.3 dBFS | Below the noise |
| 7th harmonic (7 kHz) | -121.6 dBFS | Below the noise |
| Noise between the tones | None: the error is all distortion | About -143 dBFS per bin, flat |
Rounding left distortion products as little as 27 dB below the tone. Dither removed every one of them and left the tone at its true level, in exchange for an even hiss. That trade is the whole point: hiss is steady and unrelated to the music, so it is far less noticeable than distortion that comes and goes with every note.
On loud material the cost of dither is small and predictable. In the same test on a busy signal, rounding error measured -101.1 dBFS RMS and the dithered error -96.3 dBFS RMS: TPDF dither raises the noise floor by 4.8 dB, exactly what the theory predicts.
When to dither
- Exporting a 16-bit file. A CD master, or a 16-bit WAV a store or distributor asks for. This is the case dither was made for.
- Once, at the very end. After the EQ, compression, limiter and any gain change. Nothing that changes level may come after it, or the dither stops doing its job.
- On the final file only. If a mastering engineer, a distributor or another session will process the audio again, send them a 24-bit or 32-bit float file and let the last step dither.
When to skip it
- 24-bit exports. TPDF dither at 24 bits sits about 144 dB below full scale, under the analog noise of any converter and any room. Adding it is harmless; leaving it out is also fine.
- 32-bit float files. A float export keeps roughly the resolution the DAW mixes at, so there is no meaningful word-length reduction to dither.
- Stems, bounces and files you will keep working on. Export them at 24-bit or 32-bit float and dither only the final master.
- A second time. If your DAW's export already applies dither, do not also insert a dither plugin.
Most DAWs put dither in the export or bounce dialog, sometimes listed by algorithm name. Turn it on only when the export's bit depth is lower than the project's.
TPDF or noise-shaped dither?
TPDF (triangular probability density function) dither is the standard choice: random noise spanning one step either side of each sample, which removes the distortion completely and spreads its hiss evenly across the spectrum. It is predictable and safe on any material.
Noise-shaped dither moves much of that hiss toward the very top of the spectrum, where hearing is least sensitive, so the noise is harder to hear at normal listening levels. The catch is that the shaped noise carries more high-frequency energy in total, so it belongs only on the final file. Applied twice, or followed by EQ, gain or resampling, it can make the noise easier to hear, not harder.
For most releases either is inaudible at sensible playback levels. Pick TPDF when unsure, and noise shaping for quiet, sparse material going to 16 bits.
Where dither sits in a mastering chain
Dither is the last link: tone, dynamics and density first, then a true-peak limiter to set the ceiling, then dither and the export. The dither noise is far too quiet to change the limiter's ceiling or the loudness reading. The whole chain is laid out step by step on the mastering page, and the free loudness meter checks the finished file's LUFS and true peak in your browser.
