What is dither?

Dither in plain English

Digital audio stores the height of the wave as a number, thousands of times a second. The bit depth sets how many different numbers there are to choose from. When you export a 24-bit or 32-bit session as a 16-bit file, every sample has to be rounded to one of the fewer values that 16 bits can hold, and the small difference between the true value and the rounded one is called the rounding error, or quantisation error.

How that error behaves depends on how big the sound is compared with one step. When loud music crosses thousands of steps, the error is different on every sample and acts like a faint hiss. When a sound is only a few steps tall, as at the end of a fade-out, rounding changes the shape of the wave itself, and the error turns into distortion that follows the music. Dither is a tiny amount of random noise added just before the rounding, and it keeps the error behaving like an even hiss at every level, all the way down to silence.

That sounds like a strange trade: adding noise on purpose to make something sound cleaner. It works because your ears are very good at ignoring a quiet, constant hiss and very bad at ignoring distortion that changes with the music. The lab above lets you hear both at bit depths low enough to make them obvious.

Bit depth and rounding error

Each extra bit doubles the number of steps a sample can land on. A 16-bit file has 65,536 of them; a 24-bit file has 16,777,216. The full height of the wave, from the bottom of the scale to the top at 0 dBFS, is split into that many equal steps, and every sample is stored as one of them.

Rounding to the nearest step is never more than half a step out, and on loud music the errors spread evenly between those limits. That puts the average level of the error at a fixed distance below the step size, so you can work it out for any bit depth. Every bit you take away doubles the step size and makes the error 6 dB louder.

Those are tiny numbers. At 16 bits the error is around −101 dBFS, and a loud master averages somewhere near −10 dBFS, so the error is about 90 dB below the music. The trouble starts on quiet material. A fade-out heading for silence, or the last second of a reverb tail, can sit 70 or 80 dB below full scale, and there the music is only a few steps tall. The sample rate and bit depth lesson covers the rest of what bit depth does.

What does rounding without dither sound like?

Loud music hides the problem. In step 2 of the lab the chords are rounded to 8 bits with no dither, and the error is already loud enough to hear, yet it sounds like hiss. The chords cross over a hundred steps, so where each sample lands between two steps is close to random, and the error comes out as noise without any help.

When a sound is only a few steps tall, rounding changes its shape. A smooth sine wave that climbs four steps comes out as a staircase with four treads, and a staircase is a different wave with its own harmonics. Step 6 in the lab does exactly this to a 110 Hz tone at the bit depth of a CD: the spectrum grows a row of new peaks at 330 Hz, 550 Hz and up, and you hear the clean tone turn buzzy.

At that size the error is made out of the music, so it follows the music. Step 3 turns the chords down 36 dB, which leaves them fewer than two steps tall in the 8-bit file. Press Error only and watch the Error level screen: the coloured line now rides up and down with the pale one, and the Swing figure in the coach shows by how many decibels. That's why undithered rounding on quiet sounds comes across as grit sitting on the notes, and why it's much easier to hear than its level suggests.

The worst case is a sound smaller than half a step, because every sample of it rounds to zero and it disappears from the file altogether. Try it: pick the sine, set No dither and 16 bits, then turn Level to −96 dB. The file goes silent, since the tone never gets more than half a step away from the middle.

How does dither work?

Dither adds a small random number to every sample just before it's rounded. The usual kind, TPDF (triangular probability density function) dither, is made by adding two random numbers together, so its values range up to one step either side of zero and cluster near the middle.

Say a sample should sit 30% of the way from one step to the next. Without dither, it rounds down every time. With dither, it rounds up about 30% of the time and down about 70%, so over many samples the average lands where the true value was. The individual samples are still on whole steps; the extra detail is carried by the pattern of ups and downs. That's how a tone smaller than a step can survive in a dithered file: set Level to −96 dB again and switch to Dither, and the tone comes back inside the hiss.

The cost is noise. TPDF dither makes the total error about 4.8 dB louder than plain rounding, which at 16 bits takes it from around −101 dBFS to around −96 dBFS. In exchange, the error no longer depends on the music at all. Its level holds still whatever the music is doing, which is what the flat line in the lab shows, and the Swing figure drops close to zero.

Types of dither

Rectangular (RPDF)

One random number per sample, spread evenly across one step. It's quieter than TPDF and removes the distortion, but the noise level still rises and falls a little with the music, an effect called noise modulation. Ableton Live offers it as Rectangular; for a final export, TPDF is the better choice.

Triangular (TPDF)

Two random numbers added together, as described above. It's the standard flat dither, the one the lab uses, and the safe choice for any file that might be processed again. Ableton Live calls it Triangular.

Noise-shaped

TPDF dither with a filter that moves the noise around the spectrum. It's covered in its own section next, because it behaves differently enough to need its own rules.

What is noise shaping?

Noise shaping feeds each rounding error back through a filter and subtracts it from the next few samples. The total amount of noise goes up, but the filter decides where in the spectrum it ends up. A good noise shaper pushes it down between about 1 and 5 kHz, where your hearing is most sensitive, and piles it up above 15 kHz, where it's least sensitive.

The lab's noise shaper is a five-tap filter designed for 44.1 kHz. In step 8 the coach measures its total noise as about 12 dB louder than flat dither and its noise between 1 and 5 kHz as about 18 dB quieter. Listen with Error only and the hiss sounds thin and high, like air escaping.

You'll find noise shaping in most DAWs and mastering limiters. Ableton Live and Logic Pro both include POW-r dither, where POW-r 1 is the mildest and POW-r 2 and 3 add more noise shaping (Logic Pro labels them that way in its Bounce window); Pro Tools ships POW-r as a plug-in. iZotope Ozone has its own MBIT+ dither, and the Waves L2 limiter has IDR dither with noise-shaping options.

Use it only on the final file. The shaped noise is designed for one bit depth at one sample rate. If the file is processed afterwards, by a mastering engineer, a sample rate converter or even a big EQ boost at the top, that pile of high noise gets moved or amplified and can end up more audible than flat dither would have been.

When should you dither?

Dither whenever audio is reduced to fewer bits, and do it once, as the very last step. In practice that means exporting your finished master to a 16-bit WAV, for a CD, a 16-bit download or anywhere else that asks for one.

Inside your DAW you don't need to think about it. Ableton Live, Logic Pro, FL Studio, Cubase, Reaper and Pro Tools all process audio in floating point, where the rounding error is so small that it never builds up across plug-ins and faders. Dither only comes into play at the moment the session is written out to a fixed-point file.

Streaming services accept 24-bit WAVs through most distributors and make their own streaming files from them, so a 24-bit master with no dither is a fine thing to upload. If your distributor only takes 16-bit, export that file with dither. For how those services then set the playback level, see streaming loudness normalisation, and for the loudness numbers themselves, the LUFS lesson.

How to dither your export, step by step

1. Finish the master firstThe limiter goes last on the master channel, with its ceiling set where you want the peaks. Nothing that changes the level should come after the dither.
2. Choose the bit depthExport at 24-bit or 32-bit float for your archive and for anything that will be processed again. Make the 16-bit version as a separate export.
3. Turn dither on for the 16-bit fileIn Ableton Live it's the Dither Options menu in the Export Audio/Video window, and Logic Pro has a Dithering menu in its Bounce window. Reaper puts a Dither master tickbox in the Render window.
4. Pick the typeTriangular (TPDF) if you're unsure; Logic Pro has no plain TPDF option, so use POW-r 1 there. A noise-shaped option like POW-r 2 is fine for a final 16-bit master at 44.1 kHz that nobody will edit.
5. Check that only one dither is onIf your limiter has its own dither switched on, as Ozone and the L2 can, turn off the one in the export window, or the other way round.
6. Leave the file alone afterwardsDon't normalise it or convert its sample rate. If you need to change anything, go back to the 24-bit or float master and export again.

Common dither mistakes

Dithering twice

The usual cause is a mastering limiter with its dither switched on and the DAW's export dither switched on too. Each pass adds its own noise, and if both are noise-shaped the high-frequency hiss stacks up.

Processing after dithering

A fader move, a normalise or a fade added after the dither changes every sample again, so the file has to be rounded a second time, and the first dither's noise stays in it as extra hiss. Make every fade and level change before the dither, which always comes last.

Noise-shaped dither on files that aren't final

Stems, mixes going to a mastering engineer and samples that people will pitch and process should use flat dither or none. Shaped noise is built for one exact file, and pitching it or changing its sample rate can move that pile of high hiss down to where you hear it.

Dithering a float file

Some DAWs let you tick dither on a 32-bit float export. The file holds the mix with far more precision than any converter can play, so all the dither does is add noise.

Worrying about it on a loud master

On a modern pop or electronic master the music sits about 90 dB above a 16-bit error, and the difference between dither types is very hard to hear. Turn it on and move on. Dither earns its place on quiet acoustic recordings and on anything with long fades.

Training your ears to hear dither

The quiz under the lab plays your chosen sound twice, turned down until it climbs only a step or two of an 8-bit file, once with dither and once without, with the screens covered. Listen to what the error sounds like more than how loud it is: dither gives a hiss that stays put, while undithered rounding gives a fizz that clings to the notes. There are eight rounds, and your best score is saved in your browser.

Dither and bitcrushing

A bitcrusher is rounding without dither, taken down to far fewer bits on purpose. At 4 to 8 bits the error is loud enough to become part of the sound, which is the crunchy, lo-fi character people want from it. Dither is the opposite job: keeping the same rounding inaudible. The bitcrushing lesson covers the effect, and bitcrushing vs sample rate reduction compares it with the other way of making digital audio sound rough.

Questions people ask

Keep going