Pattor says: let your ears decide whether dither at 24 bits or not

There are so many dither-related posts on this forum now that I'd like to further address some things to consider.

First of all, for those who haven't read my initial post about the issue here - please do so first, if you have any dither-questions popping your brain.

What I'd now want to provide are some thoughts of dithering a project that is exported to 24 bits. When we talk about 24 bits we talk about an integer word.

If we have recorded audio in 24 bits and we mix it in Sonar by routing things, put effects on things, eq things and adjust levels of things - then the outcome of this mix will be a floating point word. Our initial 24 bit tracks have expanded into something larger.

Within Sonar this floating point word will "shrink" to a 24 bit word when it's being sent to our soundcard. This will happen during a final step in Sonar where the internal resoultion is somewhat decreased to fit the digital pipeline in our soundcard. This digital pipeline can only transport integer words, so the floating point word must be reduced in information to fit this digital pipeline.

When this occurs there will be a slight loss of information. This information will travel along with the audio data in the digital pipeline, but this information does not contain anything musical. The good thing with playing back at 24 bits is that the level of this un-musical information (basically we're talking about distortion here) is so low that it's barely audible.

So, now as we have the ability to add a triangular dither to the output at the 24 bit level, this is what happens:

The digital distortion is masked by a very low level noise.

The noise, on the other hand, will occupy more of the information in the almost inaudible range than the distortion does in order to effectively mask the distortion. We are in a state of choosing between being hit in the face or being hit on the skull. None is really good, but the good thing is that none of the options are really very audible at a 24 bit level.


So.........


When we now talk about dither or not dither 24 bit levels (upon export or playback) I would like to point out the folowing things:

When we mix something with Sonar and have recorded in 24 bits, we will listen to a 24 bit word in our pipeline that becomes analogue and feeds our listening chain (either headphones or monitors). It will be 24 bits even though Sonar is mixing in a 32 or 64 bit mode because our soundcard can only convert digital data to analogue audio when it is a 24 bit integer wordlength.

1: If the dither is on during this mix, we won't hear the subtle effects of the distortion as the mix progresses.

2: If the dither is off, there might be a small and subtle effect of the resulting distortion that makes us take different decisions when mixing and these decisions might not be entirely based on the musical information but instead on the effect of distortion.

3: If we have mixed with dither off and switch it on during playback or add it upon export we might find that it actually did something we do not like.

When looking a the #3 scenario above there is a simple explanation. We were mixing with dither off and included the distortion in our mix. We liked the sound of our mix and the distortion was actually a part of this sound. If we then add the dither to a mix we have been working for hours, days, or months with, the way we hear the mix might alter in a tiny fashion and we might NOT LIKE IT!

That's right - WE MIGHT END UP HATING AND CURSING the effect of dither at the 24 bit level.


So the bottomline is:

If you really want to dither your 24 bit export or hear what is actually going on when you make mix decisions - you should start mixing from scratch with dither enabled. In that case only the musical infromation will pass out from Sonar, through your digital integer pipeline and finally become analogue for you to hear it.

If you start to dither after you've taken mix decisions that actually are affected by hearing distortion, you might end up dissapointed at the good things dither is supposed to do. The distortion is part of your sound. You've included it in the BIG PICTURE! And you might very well LIKE IT THAT WAY!

And I therefore recommend that if you find that you do not like what the dither does when exporting to 24 bits or when switching it on during monitoring of a mix, then you should NOT DITHER. It might sound weird, but since both the dithernoise and the digital distortion are both not musical things you have the ability to choose what fits your material.

I hope this make a point for some and that we can easier know what the good and the bad things are and how we should meet those nasty bastards in the big picture of our music.


(P.S. For clarification - the dither option in this scenario is the triangular one.)

Ditto!

ORIGINAL: krizrox

Thank you for taking the time to explain all this in laymans terms - it has helped me to understand all this considerably.

Do you really think you can hear a difference between dithered and undithered 24-bit audio?

The noise in the electronics in your DAC basically dithers the audio to roughly 20-bits or so anyway, so why would you even want to apply dither to 24-bit audio? It just adds distortion that serves no purpose. Not that you'd hear it, but still, it serves no purpose...

sinc......

The noise in the electronics in your DAC basically dithers the audio to roughly 20-bits or so anyway, so why would you even want to apply dither to 24-bit audio? It just adds distortion that serves no purpose. Not that you'd hear it, but still, it serves no purpose...

The dither is meant to mask artifacts that are the results of digital processing at higher bitdepths than the goal bitdepth. It's true that the DAC might add noise, but that is when you are hearing the playback to your ear. That noise is nothing that is included in an export, bounce or mixdown since it's rooted in a digital to analogue conversion.

The dither ain't adding distortion - it's adding noise with the purpose to mask distortion. The dither sounds like wind and the distortion sounds like breaking a glass against a concrete floor. They are both kind of disturbing, but one is more offensive.


zungle.....

It sounds like you are really implementing some heavy steps of truncation there. Since you export your project to 16 bits with no dither and then post-tweak it you are most likely amplifying the distortion. First by the dc corrector and then once more by T-racks.

Since you've carried out a mix in Sonar, eventhough the original audio was 16 bit, your mix has been in a state of a bigger bitdepth before coming back to 16 bits. Exporting this mix to 16 bits without dither is like final processing in T-racks with dither off.

You would yield much better results if you exported your mix to at least 24 bits out from Sonar, since much more of the mix information will come across to the post-editing. If you once add truncation artifacts to an audio file it can not be removed with dithering in later stages. These artifacts are smashed into your audio. Since you process it further you are only creating worse and worse sounding audio and I guess that you really are into the sound of digital graininess.

Thank You Very Much..

I will start a small Audio project tonight and practice.

Am I understanding that even though recorded @ 16/44.1, once in Sonar the audio has become part of the larger Sonar audio engine, And then when I export I am truncating those bits?

ORIGINAL: pattor

It's true that the DAC might add noise, but that is when you are hearing the playback to your ear. That noise is nothing that is included in an export, bounce or mixdown since it's rooted in a digital to analogue conversion.

I think sinc might be confusing noise at the ADC side with noise on the DAC side. But you are right. Anything added after the bit depth conversion is too late.

The dither ain't adding distortion - it's adding noise with the purpose to mask distortion.

See my other post.

UnderTow


Edit: Fixed quoting.

Hello Undertow

Dither is not about masking the truncation distortion, it is about eliminating it all together (if properly implemented). It actually allows to encode signals with amplitudes smaller than the least significant bit. Theoreticaly, if well implemented, it can increase the resolution infinitely.

In other words, always leave dither on.

UnderTow

Yes, that's right. I choosed to realte this issue to "masking" since it would perhaps make the subject somewhat easier to understand for any "newbies". That's why.

ORIGINAL: pattor


Yes, that's right. I choosed to realte this issue to "masking" since it would perhaps make the subject somewhat easier to understand for any "newbies". That's why.

Ok understood. But lets not feed the world with any more misinformation. There is allready enough as it is.

UnderTow

Pattor...

Not trying to hijack your thread (& you sound like you have it figured out, anyway)...

but...you seem to have a pretty good handle at this internal processing/precision thing...

so please have a look at my thread (http://forum.cakewalk.com/tm.asp?m=753452) & see if you can shed some light.

Thanks,

mark4man

The internal engine and effects are 32 bit floating, so you are never running what you are hearing as far as I understand. Truncating is always required.

Now here's where dither really becomes important. When you do a fade out on a mix,

My impression, being the layman that I am, was that dither is applied to mask *artifacts* that arise when going from 24 to 16 bit that occur whether there is a fade or not.
salsa

ORIGINAL: Rednroll

No. Are we reading the same thread? Always leave dither on.

(Theoreticaly you could export audio at the same bit depth as the audio engine but as Pattor correctly pointed out, you are not mixing what you hear).

UnderTow

Well, I hate to be the one to say this, but that's not correct information either. Here's the purpose of dithering. DACs are fixed point integer math devices. So if you have 24 bit DACs, when you are mixing down the floating point math will be converted to fixed point math and then the bits will be "truncated" coming out from the DAC. So in essence you are hearing what you mixed down without any dithering applied.

Yes and no. You are hearing the mix with quantisation distortion and you are not hearing the mix with the dither added which you should.

Now here's where dither really becomes important. When you do a fade out on a mix, then when the audio get's converted from floating point to fixed point the fixed point resolution gets very small where you start to be working only with 1 or 2 bits of resolution since the amplitude is so small when converted to fixed point during fade in/out situations where the amplitude becomes very small.

It isn't only on low amplitude signals. The quantization distortion affects signals of all amplitudes.

So the audio still has dynmics at this level but is only represented by 1 or 2 bits of information. What happens is that due to the dynamics the audio will sound like it's turning on and off during a fade out, because it's going from absolute digital zero and intermitantly being turned on with the representation of 1 or 2 bits. So if you listen to the fade out thru a pair of headphones the fade out will sound unnatural because there are moments of interuption where it goes to complete silence and then turns back on with those one or 2 bits of resolution that can be heard.

This is happening at the sampling rate speed. You shouldn't hear it going on and off. What you hear is quantization distortion. Also, as the quantization artifacts are not anti-alias filtered, they can have frequencies that go beyond the nyquist frequency and alias back into the audible band.

When you hear the signal going on and off is when the signal amplitude is below one bit for relatively long periods of time. Not because the signal is modulating arround one or two bits.

Anyway, this is all bad news for audio quality.

When you have a 24 bit DAC, this is hardly noticeable because your dynmics range is 144dB. So it is very difficult to hear the turning on and off of the fade out where you're only working with 1 or 2 bits of fixed point resolution. But your final destination is a 16 bit CD, which has 96 dB of dynamics range where the turning on and off of the last 2 bits of resolution in the fixed point math becomes much more obvious during a fade out.

With proper dithering you have more than 144 dB and 96 dB dynamic range respectively. You only have those figures without dither. The dither means that the noise floor is higher but we can perceive sounds well into the noise floor.

To elliminate the turning on and off that is obvious, "dither noise" is added which is usually a broadband noise with an amplitude of 1 or 2 bits. So now when this gets added to the audio signal it is less obvious that the fade out is turning on and off, since there is now the noise of the dither being added to the audio signal. So there is much less of the turning on and off of the audio signal that is noticeable during a fade out situation. So if you apply dither noise during the mixdown process, and then futher add dither noise at the mastering process, then these dual noises can add together and now you have a higher noise floor of up to 4 bits of noise.

When you add dither (could be TPDF noise, or noise shapping which isn't actually adding any noise, could be adding coloured noise, could be UV22 which seems to be a guarded secret as to what it really is) to go down to 24 bit, you are not adding it at the same amplitude as when you go down to 16 bit. The two bits peak to peak of dither (assuming TPDF) don't add up to 4 bits. Even if you add TPDF twice at the same amplitude, they don't add up to four bits as the signals are not correlated.

You still need to add dither to go to 24 bits or you end up with quantization distortion (aliased or not). By adding dither you actually increase the resolution of your signal beyond the 16 or 24 bits available. It is always good to dither your signals when reducing bit depth.

So the real summary is that you should use dither ONLY as the last process when converting to 16 bit to be transferred to the final destination of CD. This is why I say to avoid adding dither during mixdown.

Sorry but this is wrong. Only when mixing down to floating point formats (32 and 64 bit) should you not add dither.

Hopefully you're exporting at the same resolution as your DACs, therefore you ARE hearing exactly what you are mixing at.

Only when you are exporting without dither which is wrong.

Now take that 24 bit fixed point .WAV file and go to the final mastering processes. Then apply your final fade out, where you are still working with floating point math, yet still hearing the audio thru the 24bit truncation of the DAC. Then when you convert to 16 bit, then apply the dither so you don't hear the turning on/of when the audio is within an amplitude of 1 or 2 bits of fixed point resolution which is only obvious during low level amplitude signals like when you fade out to digital zero or fade in from digital zero.

This is wrong, see above.

The different types of dither are the shapes of the noise being applied to the entire audio signal which usually only has 1 or 2 bits of resolution. Thus "triangular" adds a 1 bit amplitude of a tringle wave. Then there's white noise/broadband, which random.

This is entirely incorrect. You never add a triangle wave for dithering. You add random white noise with a triangular distribution probability function. This has nothing to do with the waveform. And you need to add 2 bits (peak to peak) of TPDF noise for correct dither.

So if you are adding dither at mixdown and additionally at mastering, you are unnecessarily adding double the amount of "noise" to your audio signal, which means it can be additive noise, which means it can increase up to +6dB in level each time you add dither to your audio signal.

As stated above, you are not adding them at the same amplitude. You need to modulate the LSBs of the destination word length so you need to dither any time you reduce bit depth. Adding noise at -144 dB hardly adds up to noise added at -96 dB.

Ok understood. But lets not feed the world with any more misinformation. There is allready enough as it is.

I agree, let's try not to do that because a lot of the information in this thread is incorrect. Jesse G's summary is right on the money.

Sorry but your information is incorrect and so was Jesse G's summary. I remain with my statement about misinformation.

UnderTow

I agree with most of what you say. Just a few comments:

ORIGINAL: cGar

When ever you truncate (that means always) you have a resulting distortion that is HIGHLY correlated to the audio you are working with. Dither flatens this distortion to a white or shaped noise (depending on the dither type)

It isn't that it flattens the distortion. It actually removes it. Instead you have noise but that is preferable.

To give a working example: say you are working with singer. And you take your vocal track and filter out everthing below 1k and everything above 4k. So you only have midrange frequencies. You then truncate from 32-bit float to 16 or 24 bit. The distortion that happens will exist only in the small frequency range (1k-4k) due to the "correlation" I mentioned above.

This isn't entirely correct. The distortion is indeed correlated with the signal but try the following experiment:

In SoundForge (or similar), generate a sine wave at 4Khz. Open the Spectrum analysis and set buckets to 65,536 (top left). You will see that SF doesn't generate a pure sine wave but anyway, now reduce the bit depth to 8 bits with no dither. Refresh the Freqency analysis view. You will see that there is distortion at many different frequencies. Especially above 4 Khz. (You can also hear this distortion)

So if you must truncate before getting to the mastering phase. For example you leave the digital realm to an anolog mixer or you digital mixer only has 24-bit inputs you should dither. A small amount of white (uniform) dither noise is better than the small amount of correlated truncation distortion (regardless if you can hear it or not) you may be percieving it and that can cause you to mix/master differently!!

Entieely agreed.

UnderTow

Thanks for confusing the crap out of me, guys. I have no idea now whether to leave dither on or off during tracking and mixing. I'm not as worried about mastering, since I may hire a mastering house. Hope they know what the dither they are doing.

Sax

ORIGINAL: Saxon1066

Thanks for confusing the crap out of me, guys. I have no idea now whether to leave dither on or off during tracking and mixing. I'm not as worried about mastering, since I may hire a mastering house. Hope they know what the dither they are doing.

Sax

The analogies are never perfect but just as those who work professionally don't always agree about whether you can hear 192KHz versus 96KHz, I think you can be fairly confident that those in charge of the mastering process will be able to overcome this terrible dilemma and dither only where it's appropriate. That said, I never miss these threads!!

Richard