Dither replaces correlated quantization distortion with low-level noise when audio is reduced to a lower fixed-point precision. It does not recover detail or improve a mix. Its job is to make the unavoidable rounding error behave more benignly at very quiet levels.
The decision depends on the calculation happening at export, not a rule that every master always needs noise added. A copied 24-bit integer file has not been requantized, while processed floating-point audio rendered to 24-bit integer has. Those two exports may share a label, but they do not share the same mathematics.
Finish the musical work involved in mastering your own music without damaging the mix before making this decision. Dither belongs to delivery conversion, not tone, punch, width, or loudness. Adding it early only leaves later processors working on noise that was meant for the final boundary.
That final fixed-point conversion is the trigger. A 32-bit floating-point render kept for more processing does not need dither because it has not crossed into the shorter integer destination. An untouched 24-bit integer source copied directly to another 24-bit integer file also introduces no new quantization step.
The source label alone can mislead. A 24-bit file imported, changed by half a decibel, then exported at 24-bit has passed through higher-precision calculations and must be quantized again.
A processed session rendered from floating point to 24-bit integer is different. Although 24-bit dither sits extremely low, the conversion still discards precision. Applying suitable dither makes the error independent of the music instead of allowing low-level harmonic patterns to follow the waveform.
The audible stakes rise at 16-bit because its quantization floor is much higher than that of 24-bit audio. Quiet fades and exposed reverberation tails reveal the difference most readily. Dither turns the rough, signal-related residue into a steady noise floor that is usually less objectionable.
Sample rate and bit depth describe different properties. Changing from 96 kHz to 48 kHz does not itself create a reason to dither, but the converter performs calculations at higher internal precision. If its output is then written to a shorter fixed-point file, apply dither after conversion at that final write.
Noise shaping changes where the dither energy sits across the frequency spectrum, so less of it occupies the ear's most sensitive region. That can suit a final 16-bit distribution master that will not be processed again. It is a poor default for an intermediate file because later EQ, level changes, or sample-rate conversion can alter its carefully shaped spectrum.
Use unshaped dither when a fixed-point intermediate is unavoidable, and further work will follow. Better still, keep the intermediate at floating-point precision when every application in the path supports it. That postpones quantization until the actual delivery file is created.
Each fixed-point destination receives dither appropriate to its own word length. This does not violate the instruction to dither once because each file has one terminal quantization boundary. What creates trouble is processing a dithered derivative, reducing it again, and stacking another noise floor over the first.
If a revision is requested, reopen the high-precision source rather than equalizing or fading the reduced distribution file. Previously added dither cannot be removed cleanly, and another export may require another application. Returning to the parent prevents cumulative noise and preserves the most accurate processing path.
Confirm the required destination before selecting a dither mode or target depth. Leave dither disabled on the high-precision archive, then enable it only on the export that crosses into its final integer format. The resulting file receives one deliberate quantization step after every creative and technical calculation is complete.
The decision depends on the calculation happening at export, not a rule that every master always needs noise added. A copied 24-bit integer file has not been requantized, while processed floating-point audio rendered to 24-bit integer has. Those two exports may share a label, but they do not share the same mathematics.
Finish the musical work involved in mastering your own music without damaging the mix before making this decision. Dither belongs to delivery conversion, not tone, punch, width, or loudness. Adding it early only leaves later processors working on noise that was meant for the final boundary.
The export format decides whether dither belongs
Digital processing usually runs at 32-bit or 64-bit floating-point precision even when the source recording is 24-bit integer. Equalization, gain changes, limiting, fades, and other calculations create values that cannot be represented exactly in a shorter integer word. Writing those results to 24-bit or 16-bit PCM requires rounding or truncation.That final fixed-point conversion is the trigger. A 32-bit floating-point render kept for more processing does not need dither because it has not crossed into the shorter integer destination. An untouched 24-bit integer source copied directly to another 24-bit integer file also introduces no new quantization step.
The source label alone can mislead. A 24-bit file imported, changed by half a decibel, then exported at 24-bit has passed through higher-precision calculations and must be quantized again.
A processed session rendered from floating point to 24-bit integer is different. Although 24-bit dither sits extremely low, the conversion still discards precision. Applying suitable dither makes the error independent of the music instead of allowing low-level harmonic patterns to follow the waveform.
The audible stakes rise at 16-bit because its quantization floor is much higher than that of 24-bit audio. Quiet fades and exposed reverberation tails reveal the difference most readily. Dither turns the rough, signal-related residue into a steady noise floor that is usually less objectionable.
Processing must finish before the noise is added
Dither should follow every operation that changes sample values. That includes the last gain adjustment, fade, limiter, and sample-rate conversion. Moving it earlier allows a later calculation to reshape the noise and creates another word-length decision at export.Sample rate and bit depth describe different properties. Changing from 96 kHz to 48 kHz does not itself create a reason to dither, but the converter performs calculations at higher internal precision. If its output is then written to a shorter fixed-point file, apply dither after conversion at that final write.
Noise shaping changes where the dither energy sits across the frequency spectrum, so less of it occupies the ear's most sensitive region. That can suit a final 16-bit distribution master that will not be processed again. It is a poor default for an intermediate file because later EQ, level changes, or sample-rate conversion can alter its carefully shaped spectrum.
Use unshaped dither when a fixed-point intermediate is unavoidable, and further work will follow. Better still, keep the intermediate at floating-point precision when every application in the path supports it. That postpones quantization until the actual delivery file is created.
Each deliverable starts from the clean parent master
One release can require several bit depths without requiring a chain of conversions. Create each version directly from the approved high-precision master. A 24-bit delivery and a separate 16-bit version should be parallel exports, not a 16-bit derivative made from the already reduced 24-bit file.Each fixed-point destination receives dither appropriate to its own word length. This does not violate the instruction to dither once because each file has one terminal quantization boundary. What creates trouble is processing a dithered derivative, reducing it again, and stacking another noise floor over the first.
If a revision is requested, reopen the high-precision source rather than equalizing or fading the reduced distribution file. Previously added dither cannot be removed cleanly, and another export may require another application. Returning to the parent prevents cumulative noise and preserves the most accurate processing path.
Confirm the required destination before selecting a dither mode or target depth. Leave dither disabled on the high-precision archive, then enable it only on the export that crosses into its final integer format. The resulting file receives one deliberate quantization step after every creative and technical calculation is complete.
