dither doesnt dither

[kneelherring]

in the render dialog, tested a 16-bit file peaking at -90db, lots of truncation noise

[bobobo]

adding noise is the principle of dithering/noiseshaping

you can use it to lower low volume noise in an otherwise normal leveled setting to bring out the lowlevel parts more clearly when renderng to a lower bitlevel (f.i. 32fp to 16bit).


test : make a low level (-90 dB f.i.) sinewave (js:generator and record output

render it to 16bit with and/or without dithering and import all four

normalize all of them

Before play reduce the overall volume

undithered unnoiseshaped will be the ugliest. Lots of artefacts, you "hear" the steps.

dithered noiseshaped will be the best result , at -90 all will be covered with noise of course, but the sinewave is much more pleasing

with dither and without noiseshaping will be second best, but the noise come out too much (of cause , it's not shaped)

without dither with noiseshaping will be nearly as "ugly" as the one without all at a little lower volume (if cause, the noise is shaped and more earpleasing)

[kneelherring]

specifically, with the dither in the render dialog enabled, truncation noise is present

[bobobo]

and without dithering truncation noise is gone?

[kneelherring]

without truncates, as it should

[bobobo]

i'm too dopey to understand the problem

[bezusheist]

it looks like Reaper uses Gaussian dither...it does not perform as well as TPDF (on test signals) unless you use the noise shaping option.
TPDF FTW !

[kneelherring]

should this be addressed?

[EpicSounds]

Can you explain how we can test and verify the bug?

[bezusheist]

This is not a bug, just the type of dither used.
They should offer options for dither, like the JS dither plug in has.

[schwa]

REAPER's dither is not guassian btw. https://forum.cockos.com/showthread.php?t=23324

[bezusheist]

Huh, then maybe it is a bug because it acts just like Gaussian and not like TPDF...

[schwa]

Maybe what you are observing is just the fact that the dither is 1.5 bits which is on the lighter side -- adding less noise energy at the expense of possibly letting truncation noise through for certain highly structured or degenerate sounds. It's unlikely there is a bug in REAPER's dither.

[kneelherring]

stated as a bug simply because removal of truncation noise (distortion) is not absolute, which is the reason for dither in the first place. tpdf with a noise floor of -90.4 (16-bit) is what's needed imo

[mschnell]

Just for my own education: how do you distinguish truncation- from statistical noise in the final waveform ?

(Please don't say "by your ears". I will not accept this as a way of pin pointing bugs.)

Same question goes out for Gaussian noise vs TPDF.

-Michael

[bezusheist]

Originally Posted by schwa

Maybe what you are observing is just the fact that the dither is 1.5 bits which is on the lighter side -- adding less noise energy at the expense of possibly letting truncation noise through for certain highly structured or degenerate sounds. It's unlikely there is a bug in REAPER's dither.

Yeh, I was drawing conclusions based on one test.
I just saw something less dense than TPDF so I ASSumed it was Gauss...
Thanks for clarifying.

[kneelherring]

https://www.dropbox.com/s/3v123v84ov...led-1.png?dl=0[

-90db sine 440hz

1- flat tpdf
2- built in dither
3- truncation

[schwa]

What you're seeing is the "lightness" of the dither. A clean sine wave is a highly structured signal, so truncation noise is maximized, and 1.5 bits of noise doesn't quite mask it. This is a degenerate case.

Having said that, we could add a setting for dither bit depth in the render dialog, for people who want to add more noise.

[schwa]

Wellllll, it turns out the current render dither is 1 bit, not 1.5. We'll just change this to 1.5 bits where it was supposed to be in the first place. Thanks for the post.

[mschnell]

it would be great to see the kneelherring diagram for that

-Michael

[Bri1]

Wellllll, it turns out the current render dither is 1 bit, not 1.5.

^Oh-- ""so highly trained that he can type code that compiles correctly almost 3% of the time. "" - if you wanted a presidency--1 must increase this range to at least 5%!
How about a whole range of dithers?? for those that like to fool around-- just being dithering idjutzz..

[RJHollins]

If interested ... AIRWindows makes a whole collection of DITHER plugins.