31.3.4 Dithering
Whenever rendering audio to a lower bit depth, it is a good idea to apply dithering in order to minimize artifacts. Dithering (a kind of very low-level noise) is inherently a non-neutral procedure, but it is a necessary evil when lowering the bit resolution.

Please note that Live‘s internal signal processing is all 32-bit, so applying even a single gain change makes the resulting audio 32-bit as well — even if the original audio is 16- or 24-bit. Dither should never be applied more than once to any given audio file, so unless you are mastering and finalizing in Live, it is best to always render at 32-bit and avoid dithering altogether.

https://help.ableton.com/hc/en-us/articles/209071469-Optimizing-Windows-for-Audio

 

Didn't really say that the artefacts would be audiable, did they?

Not really adding anything to the discussion by pasting stuff from the manual. Anyway, edited xpost from https://audiosex.pro/threads/dithering-101.22921/
direct: https://audiosex.pro/threads/dithering-101.22921/#post-164792

========

I used to dither a lot , but then I've pretty much stopped altogether.

Under sensible circumstances, if you have your gain-staging correct it's:

• Inaudiable
• Useless
  
• Self-defeating

1. Lossless codecs like FLAC/WV< will produce larger files for the dither you can't hear.
2. Any audio synthesized or recorded prolly has a noise floor which acts as natural dither.
3. Lossy codecs have psychoacoustic masking. Needless to say, the relationship btw dithering & psychoacoustic compression is a complex one, and dither isn't always helpful.

See if you can tell the dithered ones from the undithered ones in this set:

http://ethanwiner.com/lullaby.zip

I couldn't. So unless you produce at -90dB , 8bit, or can hear something 60-80dB below the noise floor, don't use dither.

PS: If you can hear dither at 16bit, something's wrong.

 

I don't want to be a smart-a, and I (try) to be a composer not a mix or master eng. I want to learn so pls correct me:
IMHO:
- It absolutely doesn't matter if a DAW is 32 or 64 bit app
- It doesn't matter (for dithering) at which bit depth the DAW internally operates - all pro DAWs can go high enough
- it really doesn't matter what samples are you throwing in (except maybe if you really know they are all the same quality). Of course the quality of the samples is important for the quality, but
BUT
- it does matter at what bit depth your project is set! Even if your samples are mp3s, if you throw them in to a 44.1k 24 bit project, you'll be at 24 bit (the DAW will convert to wav aso...)
SO
- if my PROJECT is set at 44.1k 24 bit and I want to send 44.1k 24 bit file to a mastering eng. I would not dither.
- if my PROJECT is set at 24 bit and I want to master it myself to 16 bit I wold (have to) dither
- the noise is byproduct of dithering and is moved to least audible range (the dither curve). It does not fill any gaps.

 

32>24 needs dither too

 

-For sound? No it doesn't, for workflow, CPU usage, realtime performance & overall performance yes it does.
-True, all decent DAWs can go up to 32 floating points.
-Yes of course!
-Yes to the last 3 questions, to keep it simple just see it this way : Reducing bit depth = dither, always. At least from an academic point of view.

 

Last edited: May 4, 2017

While your audio file is 24 bit int (or "fixed"), the processing is 32 bit float, which only has a 23 bit mantissa. That's another reason, why dither is useless in this case.

 

Yes, the Darkkman's post is for me the best Answer to the subject, and I'm perfectly in agreement with this kind of approach.

 

would you care to expound upon that statement? where is 32 bit playback attainable for the consumer? or what data is loss when maintaining a 32 bit file?

 

Last edited: May 4, 2017

I just meant that you said "I'd only dither 16,8,4," when 24 to 16 and 32 to 24 needs it too (because the basic rule is whenever there is bit depth decrease, dithering is needed) ! 32 bit floating playback won't be "attainable" for any customer it's just that your DAW works in 32 float so you need dither when exporting 24 too; anyone corrects me If I'm wrong my dithering lessons are getting old in my head.

Maintaining 32 float doesn't lose any data but the relevance is only when you're wanting FLAC or this kind of lossless format I believe. Any others output and popular format (apart from WAV 32 ofc) then needs dithering at some point.

 

Last edited: May 4, 2017

If only people were that sensitive about transients and dynamics as they are sensitive about noise that noone can actually hear.

 

/end thread

 

You never know, do check.

Depends. Some DAWs might decode mp3 to 16bit. And where did that mp3 come from? 16bit?

yes

don't have to.

Unshaped dither doesn't move anything to inadiable ranges, but since dither is noise, it does fill gaps.

Even if they can't hear it, there might be other consequences.

=======

I've seen some incredibly dumb things done by dither. On top of the detriments mentioned above,:

Dither should be the last thing on your chain, right? If you're pushing your mix so loud, and then add dither, the dither might add clipping. After all, it's doing its job right? To change sample values? (post limiting exists. but not all have)

On silence, there's nothing to dither, so there should be no noise, right? WRONG. Some dithers (perhaps all) will add noise during silence. (autoblanking exists. but not all have) Some 'ditherers' just simply add a noise floor instead of the supposed intelligent rounding that ditherers are supposed to do , but there's no (easy) way to test. anyway, any noise floor acts like dither, right?

Shaped dither exists, but what if someone decides to sample your song & make a remix? they might strech and squeeze your audio, and suddenly your nonaudiable shaped dither is very audiable. So imo in the general case, flat or no dither is best.

Solution: don't dither (unless you're exporting to 8bit or working at -80dB RMS).

 

MP3 has no bits btw. That's why it's preferable to just encode directly from your master channel. You actually don't need dithering in that case because of the psychoacoustic algorithms it uses to encode the audio into MP3. Just encode directly from 32 bits, that's the best. It masks and compresses the most things you are not supposed to hear, like noise. Bad freaking format. So 2000. I've been using FLAC and WavPack for like 10 years now. MP3 only when I really have no other choice. It's still very useful when you want to send a quick mix to a client for assessment. But that's about it, in my world. It's also useful when you just don't care about the sound, just to have something playing no matter what, and that happens in life, too.

 

the file in question was already 24. . .

 

...if your encoder supports it (Lame does, all FhGs I testes did not). Else your signal will just get truncated to 16 bit before encoding.

 

junh1024,
Might add clipping - use ceiling while limiting, before dithering. Even dithering with aggressive noise shaping, such as Ultra from izotope mbit+ adds more than 0.01 db. So, SMALL AMOUNT OF clipping here is actual for audio limited to 0.0000 db without ceiling.

 

Last edited: May 8, 2017

Oops, that might be true. Thanks! I've been always using LAME encoder since it was released and nothing else. I use Foobar2000 for encoding or even directly from Reaper which uses LAME, too.

 

Yes, that might be the best way, but another thing I forgot to mention: Most software and nearly all hardware decoders don't decode more than 16 bit. So your MP3 player, smartphone, car, etc. will truncate it again to the upper 16 bit.