It sounds like a contradiction: to improve the quality of an audio file, you deliberately add noise. Yet dithering is one of the most solid techniques in digital audio, and understanding it prevents a very common mistake at the export stage.
The problem: truncating bits
Suppose you have a mix worked on in 32-bit float or 24-bit and you need to export it at 16-bit. The obvious solution is to drop the least significant bits. On a loud signal this is harmless, but on very quiet passages, such as a final fade-out or a reverb tail, something unpleasant happens.
When the signal occupies only a few quantization levels, the rounding error is no longer random: it follows the shape of the signal itself. It becomes correlated distortion, unwanted harmonics heard as a rough, "stepped" sound on tails. In extreme cases a reverb tail, instead of fading smoothly, breaks up into a grainy buzz before disappearing.
The solution: make the error random
Dithering means adding to the signal, before the bit reduction, very low-level noise on the order of the least significant bit. This noise randomizes the quantization error and decouples it from the signal.
The result:
- the distortion disappears and is replaced by a constant, uniform hiss that the ear tolerates far better;
- details quieter than the least significant bit remain perceptible "inside" the noise instead of being erased;
- fades descend continuously all the way to silence.
In other words, you trade an unpleasant distortion for a slight, neutral noise. It is a good trade.
Types of dither
Rectangular and triangular dither
Dithers differ in the statistical distribution of the added noise. The most widely used is TPDF (Triangular Probability Density Function), obtained by summing two uniform random noises. Its advantage is that it completely removes any dependence of the noise on the signal, at the cost of a slightly higher noise level than rectangular dither.
Noise shaping
Noise shaping goes a step further: instead of spreading the noise evenly across the spectrum, it pushes it toward frequencies where the ear is less sensitive, typically the extreme highs. The total noise may even be higher, but the perceived noise is lower.
| Option | Perceived noise | When to use it |
|---|---|---|
| No dither | Distortion on tails and fades | Never for reductions to 16-bit |
| Flat TPDF | Uniform, very low hiss | Safe, universal choice |
| Noise shaping | Lower in the midrange | Final master, only once |
Practical rules
- Dither is applied only once, at the very last step, when reducing bit depth for the final file.
- Never dither intermediate bounces: if you export stems or tracks for further processing, save them in 24-bit or 32-bit float.
- Dither comes after everything else: after the limiter, after level adjustments. Any later processing cancels its benefits.
- Don't stack noise shaping: applied several times, the noise pushed into the highs adds up and becomes audible.
- When exporting at 24-bit, dither is almost irrelevant, because quantization noise is already well below analog noise. Many apply it anyway, and it does no harm.
What about compressed formats?
MP3, AAC and Opus do not store samples at a fixed bit depth: the encoder works differently. If you export directly from a float session to a compressed codec, dither is not needed. It only makes sense when, at an intermediate step, you create a 16-bit WAV from which the compressed file will later be made.
How to hear the effect
A single experiment is enough to notice the difference:
- Take a piano note with a long tail, or a fade-out.
- Lower it by about 50 dB.
- Export it at 16-bit once without dither and once with TPDF dither.
- Re-import the files and raise the volume by 50 dB to listen.
In the undithered version the tail clearly crumbles; in the dithered version you hear a hiss, but the note fades naturally. In my work this experiment is the fastest way to explain why dither is not a detail for perfectionists.
In short
- Reducing bit depth without dither creates distortion on quiet signals.
- Dithering adds minimal noise that randomizes the error and turns it into neutral hiss.
- TPDF is the universal choice; noise shaping lowers perceived noise on the final master.
- Dither is applied only once, as the last step, and never on intermediate files.
- It is not needed when exporting straight from a float session to compressed codecs.
