Is the noise I’m hearing from bit depth reduction quantization noise?

[drissa]

hey everyone,
i have done some tests to understand bit depth reduction better and wanted to check if i’m on the right track.here’s what i did:

i took a 24-bit file in ableton and exported it to 16-bit without using dither.
then i brought the 16-bit file back into ableton, put it on a new track, inverted the polarity, and did a null test against the original 24-bit file.
as expected, the files canceled each other out, but when i boosted the master by about 36 dB, i heard some noise.
i ran the test again, but this time i boosted the original 24-bit file by 10 dB before exporting to 16-bit (still no dither). when i did the null test again, the noise i heard was about the same level as before

so now i’m wondering:
Is the noise I’m hearing actually quantization noise from reducing the bit depth?
And when someone applies dither afterward, is its main purpose to mask/smooth out that quantization noise I faced? Is that essentially the whole point of dithering?

[BertKoor]

If the residual noise you get (from summing) is at about -94dB FS, then that's the truncation and dithering noise. Its level is roughly halve the least significant bit of the target bit depth.
You can try without dithering, then it should be only noise from truncation.

A better illustration is to not use 16 bit target but 8. Then the difference between with and without dither should be audible. Like it is visible here:

Left to right: original, plain truncated, with dither. (from https://en.wikipedia.org/wiki/Dither )

Best look at the white-grey gradient in the neck. By adding a bit of noise the impression of a smooth gradient is preserved despite truncation.

[lfm]

And when someone applies dither afterward, is its main purpose to mask/smooth out that quantization noise I faced? Is that essentially the whole point of dithering?

Well, you do dither while having 24-bit, once 16-bit it has no purpose, damage done one could say.
- I guess you mean "last part of the rendering process"
- if to export like you described keep 24-bit and import for final dithering and mastering stuff.

One explanation I like is this
- in 24-bit samples there are 256 levels for each single level in a 16-bit file
- truncation of these 256 levels may create patterns of the content converted, since done on the same numeric spot every time, for every sample.
- our brains are always looking for patterns and make various pitches from it

What dither do adding noise is create a randomness to the very spot on each sample where it is truncated.
- so no clear unwanted pattern is created that can be perceived by our brain

Then there are various algos how to create the noise to not inflict on the content in the samples so you actually here noise added.

My view is that it's not to mask quantization noise, but to reduce it.
- less artifacts are created in bit depth reduction, in theory
- why I say "theory" is that it depends on content in samples and what noise you add and which level etc

So that's why there are different algos that work better for different material.

[Jac459]

Interesting stuffs!
Thanks for sharing.