I have been reading a lot of articles explaining dither (because I didn't understand what that setting did in my sequencer program) and I now have a good understand about what dither is but none of the articles have cleared this up?
When recording at 24-bit does dither need to be used if the dynamic range is small and the peak levels around -6Bd?
Secondly, just thought of this whilst typing, can anyone tell what the term "32-bit float" means?
Thanks everyone.
Comments
Thanks for the feedback, I have a lot of time for reading at the
Thanks for the feedback, I have a lot of time for reading at the moment because my vintage monitor system has..erm..stopped working...well the right speaker has and mixing with one speaker is easier said than done. I know what the problem is the volume potentiometer right channel have broken but I have not got a replacement, so reading it is. Thanks
I have to slightly disagree on audiowkstation's comment about di
I have to slightly disagree on audiowkstation's comment about dither.
Certainly true, dither is used in word length conversions.
But, what dither actually is, is the addition of a bit of "noise" in the least significant bits (LSB) of the sample word. The PROBLEM with digital is, that at the bottom, you have to either be ON or you have to be OFF, there's no SORTOFONSORTOFNOT bit. what happens, is that the threshold where the last bit turns on or off results in a small amount of error in the conversion. This error manifests itself as distortion, or noise. Also called "quantization error/noise" Without dither, this noise ends up being highly correlated with the signal iteslf, which as it turns out, makes it ugly and noticeable. Adding that bit of dither noise at the very lowest levels, makes the quantization error UNCORRELATED with the signal, which, as it turns out, sounds much nicer to the ear.
So, having said that, dither is present at ANY conversion into digital, whether at the original conversion from analog, or at subsequent conversions.
A digital conversion happens every time you mess with the file in any way...such as when you move the "fader" in your DAW, or apply a plugin, or mix.
So, dither comes into play ALL THE TIME, not just when you're doing word length conversions (i.e. 24 bit to 16 bit)
Now, on to 32 bit floating point:
floating point math is simple to grasp...its when the number of digits to the right of the decimal point can change. For example, if I divide 10 into 5, I get a result of "0.5", or one place to the right of the decimal point. if I divide 10 into 6, then I get "0.66666666666...forever....".
In a floating point system, that uses a digital word that is 32 bits wide, the biggest (actually, smallest) number that can be represented is 2^32 (two to the thirty-second power), AFTER the decimal point. So, our example above becomes "0.66666666667"
In fixed point math, the number of places to the right of the decimal point stays the same. So, in our above example, divide 10 into 6 becomes, say "0.6667"
The long and short of it? that irrational numbers are much more accurately represented in floating point.
and again, 32 bit simply means that the word length is 32 bits.
My guess is that you're referring to something like the internal mix precision of a DAW...where it is performing all its calculations in 32 bit floating point math, then converting back to the original, say, 24 bit word length of the file being processed (using dither!).
With 24 bit files, there are many who argue that the inherent system noise level is adequate to perform the dithering, but there are many who disagree...so take your pick. The underlying hypothesis is that randomized quantization noise (as in, with dither) is not NEARLY as cumulative as correlated quantization noise (without dither), over the course of a significant number of operations.
dwoz
Thanks you are right, I did get the term "32-bit float" of a DXi
Thanks you are right, I did get the term "32-bit float" of a DXi plug-in and you explanation fits it perfectly so I am guessing your right there.
my average noise when recording is -50dB which i think is more than adequate for the the A-D conversion but dither will be needed when mixing into lower bits and when running signals into plug-ins with a higher float than their output. Have I got that right?
Generally (and there are always exceptions) there is not much us
Generally (and there are always exceptions) there is not much use in adding dithering except at the very last step going from 24 bits (or float) internal resolution down to 16 bit for a CD.
Very few, if any, recordings are truly 24 bit signal anyway, even the very best AD converters come in at 20 bits or less of true signal. The rest is some kind of noise and in my experience that takes the role of the dithering noise. So no need to add even more noise, even if it is at -140dBFS level or so.
By the way, if you have only one monitor, try mixing everything mono. It is a trick used by quite a few engineers: mix til it sounds good in mono, positioning in stereo afterwards. If you can make it sound good in mono, it will often sound better in stereo, the other way round is not always true.
Gunnar
My plugs work in 24bit mode and what I am working on is 24bit an
My plugs work in 24bit mode and what I am working on is 24bit and internally, everything stays in 24bit until I drop to 16 bit.
I think the moral of all of this is to KNOW your system and how it works. If each channel of a DAW is doing word length conversion while using plug ins, you need to know. Dither on top of dither on top of dither is not a good scene. Know your flow chart and signal path.
dwoz wrote: This error manifests itself as distortion, or noise
dwoz wrote: This error manifests itself as distortion, or noise. Also called "quantization error/noise" Without dither, this noise ends up being highly correlated with the signal iteslf, which as it turns out, makes it ugly and noticeable. Adding that bit of dither noise at the very lowest levels, makes the quantization error UNCORRELATED with the signal, which, as it turns out, sounds much nicer to the ear.
I don't think that is quite correct. You imply that dither noise merely disguises the quantisation distortion in a way that is subjectively more pleasant: my understanding is that dither actually preserves information that would otherwise have been lost below the LSB level. ie: it reduces quantisation error, it doesn't just disguise it..
IIRs wrote: [quote=dwoz] This error manifests itself as distorti
IIRs wrote: [quote=dwoz] This error manifests itself as distortion, or noise. Also called "quantization error/noise" Without dither, this noise ends up being highly correlated with the signal iteslf, which as it turns out, makes it ugly and noticeable. Adding that bit of dither noise at the very lowest levels, makes the quantization error UNCORRELATED with the signal, which, as it turns out, sounds much nicer to the ear.
I don't think that is quite correct. You imply that dither noise merely disguises the quantisation distortion in a way that is subjectively more pleasant: my understanding is that dither actually preserves information that would otherwise have been lost below the LSB level. ie: it reduces quantisation error, it doesn't just disguise it..
close, but not quite! What you're referring to is that valid signal can actually be discerned below the level of the quantization noise. This is not quite the same thing as saying that the signal is preserved below LSB.
Also, yes indeed, as you say, the quantization error is decreased, because statistically ANY application of dither noise will result in less quantization error WRT the input (i.e. "information") signal. (or rather, what I should say, is that the quantization error of information signal + dither will never be worse than quantization error of information signal alone) But, the main benefit is the un-correlation (is that a real word?) of the quantization noise to the information signal.
and, finally, audiowkstation: what do you suppose is the internal precision of some of those plugs you're talking about? Some of 'em are 32 bit, some are 48 bit...as such, when they output, they either have to truncate back down to 24, or dither down to 24.
AS you note, the more stuff you do, the more you tangle up those lower bits...this is one of the ugly sisters hidden in the closet of digital processing.
Digital copies are pristine...digital processing is anything but.
As has been noted, this may be so much tilting at windmills, but it may also be the difference between "whack" and "sack".
ain't this stuff FUN!!!!!!
dwoz
dwoz wrote: close, but not quite! What you're referring to is
dwoz wrote:
close, but not quite! What you're referring to is that valid signal can actually be discerned below the level of the quantization noise. This is not quite the same thing as saying that the signal is preserved below LSB.
Isn't it? I don't see the distinction.. :?
dwoz wrote: Now, on to 32 bit floating point: floating point m
dwoz wrote:
Now, on to 32 bit floating point:floating point math is simple to grasp...its when the number of digits to the right of the decimal point can change. For example, if I divide 10 into 5, I get a result of "0.5", or one place to the right of the decimal point. if I divide 10 into 6, then I get "0.66666666666...forever....".
In a floating point system, that uses a digital word that is 32 bits wide, the biggest (actually, smallest) number that can be represented is 2^32 (two to the thirty-second power), AFTER the decimal point. So, our example above becomes "0.66666666667"
In fixed point math, the number of places to the right of the decimal point stays the same. So, in our above example, divide 10 into 6 becomes, say "0.6667"
The long and short of it? that irrational numbers are much more accurately represented in floating point.
and again, 32 bit simply means that the word length is 32 bits.
Check that to say dividing 3 into 2 (2/3) and Dwoz will be absolutely correct.
Kids, there's bunches of dither going on all over your DAWs, trust us on this one.
~S
IIRs wrote: [quote=dwoz] close, but not quite! What you're ref
IIRs wrote: [quote=dwoz]
close, but not quite! What you're referring to is that valid signal can actually be discerned below the level of the quantization noise. This is not quite the same thing as saying that the signal is preserved below LSB.
Isn't it? I don't see the distinction.. :?
my bad, I misspoke.
"valid signal can actually be discerned below the level of the dither noise".
that's better.
dwoz
awrrright, you nailed me, mr "get yer terminology straight". here's a cookie :-)
dwoz
dwoz is correct. Possibly some are confused by the following
dwoz is correct.
Possibly some are confused by the following common misconception: that the terms "dither" and "word-length reduction" are synonymous. They are not. Dither is indeed part of the process used in reducing word length. But it is used in many other digital applications. Already mentioned was how distortion from a/d quantization error is tamed with dither. Anywhere there is digital processing, there is a high likelihood that some dithering is happening under the hood.
Depends on how you look at things. If you believe that analog
Depends on how you look at things.
If you believe that analog audio has "infinite" wordlength, then by all means, yes.
If you believe that analog audio has "undefined" wordlength, then you've got a bit of a semantic conundrum.
Does divide-by-zero result in an infinite result, or an undefined one? is there really any difference?
dwoz
IIRs wrote: ... there is information below the LSB level that ca
IIRs wrote: ... there is information below the LSB level that can be partially preserved using dither.
I am uncomfortable with this statement. To my understanding, dither masks inaccuracies at the LSB. Something that is masked or disguised can hardly be said to have been preserved. This sounds like the propaganda that claims that certain dithering algorithms will "preserve 24bit resolution in the 16bit domain". Marketing rhetoric, IMO.
Don't take my word for it: Bob Katz wrote: although the quant
Don't take my word for it:
Bob Katz wrote: although the quantization steps of a 16-bit word can only theoretically encode 96 dB of range, with dither, there is an audible dynamic range of up to 115 dB!
http://www.digido.com./portal/pmodule_id=11/pmdmode=fullscreen/pageadder_page_id=27/
:)
Zilla wrote: [quote=IIRs]... there is information below the LSB
Zilla wrote: [quote=IIRs]... there is information below the LSB level that can be partially preserved using dither.
I am uncomfortable with this statement. To my understanding, dither masks inaccuracies at the LSB.
Dither RADOMIZES inaccuracies at the LSB. Subtle but salient difference.
Its very much like what happens with analog tape, where you can hear program meterial below the noise floor. In "theory", that shouldn't be audible, but it is.
Again, the typical listener/practioner of digital audio has no more than a passing interest in this, but it isn't to be ignored if you're producing masters.
dwoz
dwoz wrote: Dither RADOMIZES inaccuracies at the LSB. Subtle b
dwoz wrote:
Dither RADOMIZES inaccuracies at the LSB. Subtle but salient difference.
The dither may seem random, but actually it is carefully calculated to average out to zero. The signal below the LSB biases the dither, so that it no longer averages out to zero.. this bias contains useful information, and explains why "you can hear program meterial below the noise floor"
Bob Katz wrote: .. the characterisation of the system with dither on is transformed from completely deterministic to one of statistical probability. The periodic alteration of the LSB between the states of 0 and 1 results in encoding a source value that is smaller than the LSB
Hello everyone, isnt dither great I have just had an idea: When
Hello everyone, isnt dither great I have just had an idea:
When a 24-bit digital track is dithered down to 16-bit the added dither is know because it is created by an algorithm and therefore can be recorded and stored as a wave file. The dither is created to a 24-bit resolution. If this dither was recorded and then converted down to 16-bit WITHOUT dither it would be the same dither with half the resolution and almost resemble the 24-bit dither signal. If the phase is then inverted and the file combined with the 16-bit track this would remove the dither noise from the signal and leavel the recovered elements saved by dithering.
I am just guessing this I am probably not right just a through I had.
But the dither signal much be more that 1 Bit in amplitude becau
But the dither signal much be more that 1 Bit in amplitude because otherwise its a square wave not noise, so the content recovered by the dither, if the dither was removed it would leave you with a square wave of the content recovered, or am I still wrong, your probably right, just checking
annannienann wrote: Hello everyone, isnt dither great I have jus
annannienann wrote: Hello everyone, isnt dither great I have just had an idea:
When a 24-bit digital track is dithered down to 16-bit the added dither is know because it is created by an algorithm and therefore can be recorded and stored as a wave file. The dither is created to a 24-bit resolution. If this dither was recorded and then converted down to 16-bit WITHOUT dither it would be the same dither with half the resolution and almost resemble the 24-bit dither signal. If the phase is then inverted and the file combined with the 16-bit track this would remove the dither noise from the signal and leavel the recovered elements saved by dithering.
I am just guessing this I am probably not right just a through I had.
What in the holy hell did you just say?
IIRs wrote: Don't take my word for it: [quote=Bob Katz] althou
IIRs wrote: Don't take my word for it:
[quote=Bob Katz] although the quantization steps of a 16-bit word can only theoretically encode 96 dB of range, with dither, there is an audible dynamic range of up to 115 dB!
I have heard/read all types of mathematical explanations regarding this. All seem perfectly cogent. Let me restate my position by way of an example:
I play back a 24bit recording of a solo piano. It sounds pleasing, much like the mic feed was. Now I listen to the WLR/dithered down 16bit version of the same recording. Hmmm... funny, in comparison it sounds a little like a sampled piano. Back to the 24bit file: ahhh ya, I like that. Back to the 16bit: no, it definitely sounds more like its coming from a korg or emu.
So the point I am trying to make: the 16bit version does not convince my ear that dithering preserved the resolution I had with 24bits. I can agree that dither can preserve a small degree of low level information. I also agree that, to some extent, we can "hear through the noise floor". But achieving eight extra bits of resolution from one dithered bit is beyond my listening experience.
Of course 24 bit audio sounds better than 16! I don't think anyo
Of course 24 bit audio sounds better than 16! I don't think anyone claims their dither algo can preserve all the extra resolution, but some is better than none!
Try truncating your piano recording to 16 bit with no dither, and compare that to a version using a good dither algo.. that would be a fairer test (Assuming that 16-bits is not negotiable, and our choice is between different flavours of dither, or none at all)
IIRs wrote: ...I don't think anyone claims their dither algo can
IIRs wrote: ...I don't think anyone claims their dither algo can preserve all the extra resolution, but some is better than none!
Show me an advertisement for a dither algorithm that claims only partial preservation! This is what I meant by marketing rhetoric, the kind that uninformed newbies are misguided by. But I agree with you, some is better than none.
;) I ran a snare sample through a church impulse to generate a
;)
I ran a snare sample through a church impulse to generate a long smooth reverb tail. I then dropped the gain by approx 50dB, and rendered a 24-bit file, a 16-bit file with no dither, another using the free MDA dither plug, and a final one using the MegaBitMax dither algo from the Ozone "mastering" plug. I normalised the results in Soundforge so you can listen to them at sensible monitor levels!
Dither is used when converting from 24 bit to 16 bit. 24bit has
Dither is used when converting from 24 bit to 16 bit. 24bit has much more dynamic range than 16 bit. 32bit float is a "type" of word length. You have fixed point and floating point. "Float" is obviously, floating point.
digido.com has some wonderful articles to read about digital. It is a great place to absorb knowledge as well as the "rane" glossary at http://www.rane.com
or better yet...
http://www.rane.com/digi-dic.html
If you go over that glossary at the rane site, you will have forgotten more than quite a few people in this business know!
Have fun with it!