Dither is a small amount of deliberately added random noise that is applied to a digital audio signal before its bit depth is reduced, for example from 24-bit to 16-bit. In digital audio it turns the correlated, signal-dependent error of quantization into a steady, signal-independent noise floor, which the ear tolerates far better than distortion.
Dither belongs to the final steps of mastering and file conversion. Without it, quiet passages such as fade-outs, reverb tails and soft piano notes can acquire grainy distortion and harmonics that were never in the original recording.
Why quantization needs dither
Quantization maps each sample to the nearest available level. With 16-bit PCM, there are 65,536 levels and the step size is fixed. For loud, complex signals the rounding error looks like random noise, but for very quiet or simple signals the error follows the waveform and becomes harmonic distortion and, in extreme cases, a gated, buzzing sound as the signal falls below one step.
Adding noise of about one to two least significant bits (LSB) in amplitude before rounding randomizes which level each sample lands on. Averaged over time, the output then follows the original signal even below one LSB, so very quiet details remain audible as a tone within the noise rather than disappearing or distorting.
The theoretical noise floor of an undithered quantizer is about 6.02 dB per bit plus 1.76 dB below a full-scale sine wave. Proper TPDF dither raises the noise floor by about 4.8 dB, a small price for removing distortion and noise modulation.
Types of dither noise
Dither types differ in their probability distribution and in their spectrum. The distribution determines whether the error becomes truly independent of the signal; the spectrum determines where the noise is heard.
| Dither type | Amplitude | Effect on error | Typical use |
|---|---|---|---|
| RPDF (rectangular) | ±0.5 LSB | Removes distortion, noise still modulates | Rarely used alone |
| TPDF (triangular) | ±1 LSB | Error mean and power independent of signal | Default for general conversion |
| Gaussian | Similar power to TPDF | Comparable to TPDF, slightly more noise | Analog-like behavior |
| High-pass TPDF | ±1 LSB, tilted spectrum | Less audible low-frequency noise | Mastering for 16-bit |
| Noise-shaped dither | TPDF plus error feedback | Noise moved to less audible frequencies | CD mastering of quiet material |
TPDF dither is the most widely recommended general choice, because it eliminates both distortion and noise modulation, the audible pumping of the noise floor with the music. Noise-shaped variants lower the perceived noise further but put more energy at high frequencies.
Noise shaping and perceived noise
Noise shaping feeds the quantization error back through a filter so that the total noise is redistributed in frequency. The noise is reduced in the 2-5 kHz range, where hearing is most sensitive, and increased near the top of the band, where it is much less audible.
Aggressive noise shaping can lower the perceived noise floor by roughly 10-20 dB for a 16-bit file, at the cost of a higher total noise power. Because the noise is concentrated near 15-22 kHz, it must never be processed again: any later gain change, sample rate conversion or lossy encoding may expose or destroy it.
| Choice | Perceived noise | Total noise power | Robustness to later processing |
|---|---|---|---|
| Flat TPDF | Higher | Lower | High, safe for intermediate files |
| Mild noise shaping | Lower | Moderate | Moderate |
| Strong noise shaping | Lowest | Highest | Low, final delivery only |
When to apply dither and when not
The rule is simple: dither once, at the last bit-depth reduction. Modern digital audio workstations process audio in 32-bit or 64-bit floating point, so internal processing rarely needs dither, but every export to a fixed-point format with fewer bits does.
| Situation | Dither | Reason |
|---|---|---|
| 24-bit mix to 16-bit master | Yes, TPDF or noise-shaped | Quantization error is audible at 16 bits |
| Float mix to 24-bit file | Optional, TPDF | Error lies near the analog noise floor |
| Exporting 32-bit float | No | No reduction in precision |
| File for lossy encoding | Prefer flat TPDF or deliver 24-bit | Encoder alters high-frequency noise |
| Intermediate stems | No, keep float or 24-bit | Avoid stacking dither |
Dithering multiple times accumulates noise. Each reduction adds another layer, so stems, premasters and archive files should stay at 24-bit or float, and dither should be applied only to the final 16-bit or other reduced-precision delivery.
Dither compared with related concepts
Dither interacts closely with the other basics of digital audio, and its role is easiest to understand alongside them.
- Quantization creates the error that dither makes benign.
- Bit depth sets the step size and thus how much dither noise is needed.
- Sample rate limits the band in which noise shaping can move noise; the Nyquist frequency is its ceiling.
- PCM is the format whose fixed-point samples require dither when shortened.
- Truncation without dither simply cuts bits and causes distortion and a slight DC offset.
The same principle appears in images and video, where dither prevents banding in gradients. In audio, the audible symptom is not bands but distortion and grain in quiet passages.
In simple terms
Digital audio is sometimes shortened to fewer bits. Very quiet sounds must then be rounded to a coarse set of steps. That rounding can sound gritty and harsh. Dither is a very faint hiss added just before the rounding. It hides the steps and keeps quiet sounds smooth.
It is like smoothing a staircase with a thin layer of sand. The steps are still there, but you no longer stumble on them. The added hiss is usually far below anything a listener notices. The harsh distortion it prevents would be clearly audible.
A practical example
A mastering engineer prepares a quiet acoustic album, mixed at 24-bit and 96 kHz, for CD release at 16-bit and 44.1 kHz. Several tracks end with long reverb fades that drop to very low levels.
All processing runs in floating point. As the very last step, after sample rate conversion and limiting, the engineer applies noise-shaped TPDF dither during the export to 16-bit, while the high-resolution version for download is exported at 24-bit with flat TPDF dither.
Comparing undithered and dithered 16-bit versions with the fade-outs amplified, the undithered file shows audible grain and distortion in the tails, while the dithered file decays smoothly into a quiet, even hiss. The CD master is approved with clean, natural fade-outs.