Everything you do in a 32-bit float session is reversible in a way that the final export is not. The last step of your process, reducing from 32-bit float to 16-bit, is the only genuinely destructive operation in most workflows, and it is the one most often skipped.

What truncation does

A 32-bit float file has an enormous amount of headroom above and below the nominal range. Reducing it to 16-bit means rounding every sample to one of 65536 possible values. If you simply round, the error between the original value and the nearest representable one is not random. It is correlated with the signal, which means it behaves like a low-level distortion of the waveform rather than like noise.

Correlated error is audibly worse than random error. It adds harmonics that were not in the recording, and because it is correlated with the signal it does not average out over time the way uncorrelated noise does. The result sounds dirty rather than quiet.

This is why an un-dithered export is not merely a slightly worse file. It is a file containing distortion that was never in your mix.

What dither does instead

Dither adds a small amount of noise deliberately, before the rounding step. The noise makes the rounding error random rather than correlated, which converts the distortion into broadband noise that is far less objectionable.

The amount of dither required depends on how many bits you are reducing by. Reducing 32-bit float to 16-bit requires roughly 16 bits of dither, because each bit discarded needs about one bit of noise added to decorrelate the error. The result sits around 96dB below full scale, which is well below the noise floor of the best converters and, critically, below the noise floor of whatever plays the file back.

Shaped dither

Straight dither distributes noise evenly across the spectrum, which places energy at frequencies the ear is most sensitive to. Shaped dither pushes the noise away from those frequencies, toward the very top or the very bottom of the audible range, where it is less noticeable.

Shaped dither is strictly better than unshaped for the same bit depth, and most mastering plugins include it. The only real consideration is that shaped dither should be used once, at the very end. Applying it mid-process, or dithering to 16-bit and then processing further, defeats the purpose entirely because subsequent processing will treat the noise as part of the signal.

Lossy formats are a separate problem

Dithering addresses bit depth. It does nothing about MP3 or AAC, which are lossy formats that discard information regardless of bit depth. Converting a dithered 16-bit master to MP3 discards a great deal more than the dither addressed.

The practical rule is to dither to the bit depth of your delivery format, not to 16-bit as an intermediate. If you are delivering 24-bit, dither to 24-bit or skip it. If you are delivering 16-bit, dither to 16-bit. If you are delivering MP3 from a 24-bit master, either skip the dither entirely or set it to the equivalent of 16 bits, since MP3 encoders work at 16-bit internally.

Getting the export right

1

Export from a 32-bit float session at the highest practical quality. The intermediate file should be 32-bit float or 24-bit. This is the version you keep as the archival master.

2

Never modify a 16-bit file further. Every processing pass on a 16-bit file works on already-quantised data, and the errors compound. If you need to change something, go back to the float session.

3

Apply dither exactly once, at export. Not in the mastering chain, not on a print, and not on an intermediate. Once, at the point where bit depth actually reduces.

4

Export a matching MP3 or AAC from the archival master, not from the 16-bit file. Running lossy encoding on already-dithered 16-bit data gives a worse result than the alternative, because the codec is working with quantisation noise as though it were signal.

What to deliver and what to skip

FormatDither settingNotes
32-bit float WAVNoneArchival master, never modify downstream
24-bit WAV24-bit or noneStandard professional delivery
16-bit WAV16-bit, shapedOnly when the format requires it
MP3 / AAC16-bit or noneEncode from the float or 24-bit master
Processing on 16-bitN/ANever, errors compound with every pass

Retuning and the final export

A retuning plugin is a processing pass, so it belongs in the float session rather than on a dithered file. If you use 432Hz MASTER or RASTA 528Hz on the master bus, run it before the export stage, never after it. Running it on a 16-bit print means the plugin is reconstructing a signal that already carries quantisation error, and it will faithfully reproduce that error in its output.

The bottom line

Dither is a one-second setting that determines whether your master is clean or quietly distorted. It goes on once, at the end, and it goes on every time you reduce bit depth.

Frequently asked questions

Do I need to dither when exporting?

Yes, if you are reducing bit depth. Dither converts the correlated truncation error into uncorrelated noise, which is far less audible. Without it, the error behaves as low-level distortion of the waveform rather than as a noise floor.

How much dither should I use?

About one bit of dither for every bit you discard. Reducing to 16-bit needs roughly 16 bits of dither, which sits about 96dB below full scale. Shaped dither distributes that noise toward frequencies where the ear is less sensitive.

Does dither make my file noisier?

It adds a small amount of noise deliberately, but the result sounds cleaner than the alternative. Without dither, truncation produces correlated error that behaves as distortion. With it, you get a steady noise floor below the level of your converter, which is the correct trade.

Can I edit a 16-bit file in my DAW?

Technically yes, but it is a bad idea. The file already contains quantisation error, and every processing pass works on that error and adds more. Return to the float session if you need to make changes, and re-export with dither.

More from the atixUniverse blog