Only when the bit depth goes down, and only once, as the last operation in the chain. A 32-bit float session rendered to a 24-bit WAV is a reduction, so it gets dither. A 24-bit master rendered to 16 bits for a CD or a DDP image is a bigger reduction, so it certainly gets dither. A file that keeps its bit depth gets nothing. That is the whole rule, and the rest of this piece is why it holds, where noise shaping earns its place, and which deliveries in 2026 still drop bits at all.
What does dither actually fix?
Rounding. Every sample in a digital file is stored at a fixed word length, and when a 32-bit value has to fit into 16 bits, the low-level detail has to be rounded away. Plain truncation rounds in a way that follows the signal, and error that follows the signal is distortion. It is quantization distortion, it lives at roughly the level of the last bit, around -96 dBFS in a 16-bit file, and it sounds wrong in a specific way: fades turn grainy, reverb tails break up into a fizzy texture, a solo piano decay develops a buzz that was never played.
Dither replaces that correlated error with an uncorrelated one. Adding one or two least-significant bits of random noise before the rounding step breaks the relationship between the signal and the error, and the result is a steady, benign noise floor instead of distortion. The trade is noise for truth: below the last bit the file now behaves like analog tape at a very quiet level, and detail quieter than the noise floor remains audible through it, the same way a fade stays musical on tape. At 24 bits the same mechanism operates near -144 dBFS, which is below the self-noise of any converter, preamp or room on earth, so dither at 24 bits is a correctness decision rather than an audible one.
Which deliveries still drop bit depth in 2026?
Fewer than people assume, which is why the question has a short answer now. Distributors take 24-bit masters. Apple asks for exactly that in its Apple Digital Masters specification: 24-bit sources at the highest native sample rate, at least 1 dB of headroom against encoder clipping, and their own encoding chain working from the 24-bit file directly. The document is explicit that encoding from the full-resolution source removes the need for added dither at their end. The 24-bit WAV I print as the delivery package is the deliverable; the lossy versions are made from it, not by me.
The reductions that remain are concrete. A CD or a DDP image is 16-bit 44.1 kHz by format definition, so any run of physical discs means a dithered 16-bit render. Some sync and broadcast houses still request 16-bit WAVs for their libraries. A client burning reference discs for a car gets the same treatment. And inside the production itself, the render from the 32-bit float mix bus to a 24-bit fixed-point file is a reduction too, which is why the dither switch in the export dialog exists at all. Streaming loudness and codecs are a separate subject with their own rules, and I covered that chain in how platform normalization treats a master.
Does dither go before or after sample rate conversion?
After, always. Sample rate conversion is filtering and interpolation, and its arithmetic raises the word length of every sample it touches, so dither applied before SRC is destroyed by it. The working order in my room: all processing at the session rate, then SRC to the target rate if the delivery needs one, then dither as the final operation at the final bit depth, then nothing else. A 96 kHz master headed to a 44.1 kHz 16-bit CD gets converted first and dithered second, in that order, every time.
The one-pass rule matters as much as the order. Dither applied twice stacks two noise floors, and dither applied to a file that then gets processed again turns shaped noise into unshaped problems. This is also why the question "which plugin last" has a boring answer: the dither stage is the last plugin, after the limiter, and if the limiter is also printing the final file, its built-in dither does the job and nothing else should run downstream. When an album master comes back for one more edit, I go back to the undithered print, make the change, and render again, a discipline that lives in the same folder habits as session versioning and recall.
What is noise shaping, and when do I leave it off?
Flat TPDF dither spreads its noise evenly across the spectrum. Noise shaping moves that energy where hearing is least sensitive, pushing most of it above roughly 15 kHz, so the audible band sits several dB quieter for the same total noise power. Algorithms like POW-r and the shaped modes in mastering limiters are variations on that curve. On a final 16-bit master it is a free improvement: the floor around a fade sounds lower, and the file plays back the same everywhere.
I leave shaping off in two cases. A file that will be processed again, resampled, or summed with others gets flat dither or none, because gain changes and SRC fold that high-frequency energy back where it was never meant to sit. And a 24-bit render gets flat dither on principle, since shaping noise that already sits at -144 dBFS optimizes something nobody can measure in a room. The shaped option is for the last render of the smallest word length, once, which in practice means the 16-bit masters of an album run prepared for disc.
How do you actually hear what dither does?
With a null test and a gain stage, in about five minutes. Print the same master twice at 16 bits, once truncated and once dithered. Load both against the 24-bit original, invert polarity, and listen to what remains with 40 to 60 dB of makeup gain. The truncated difference signal follows the music, a crackling shadow of the mix that rides the fades, which is distortion by definition. The dithered difference is steady hiss that ignores the music entirely. Then listen to the end of the last chorus on each 16-bit file at a loud monitoring level: the truncated tail fizzes, the dithered tail just gets quieter. Anyone mastering their own productions should run this test once, because after hearing it the export settings stop being folklore.
Frequently asked
Should I dither when exporting a 24-bit master? If the session runs at 32-bit float, the export is a reduction, so dither at 24 bits is technically correct and costs nothing. It sits near -144 dBFS, below any converter's own noise, so treat it as hygiene rather than sound.
Does dither go before or after sample rate conversion? After. SRC arithmetic raises the word length again and undoes dither applied before it. Process, convert, then dither last at the final bit depth.
Should stems or mix bounces be dithered? No. Keep working files at 24-bit or 32-bit float and dither once, on the final deliverable. Stacked dither is stacked noise with no benefit.
Before your next release ships
Send me the master and the list of destinations. I will print the delivery set, 24-bit at the native rate for distribution and a dithered 16-bit pair if discs are in the plan, and verify the renders null against the source above the noise floor. Credits are at /credits, background at /about.