Topic: Dithering facts and fiction

[mmmatt]

Ok... first off I would like to say I claim to be NO expert on dithering, but I thought I would open up a discussion on it.  I was looking through a few of the posts in this thread relating to dithering.  There were a few places where it was stated or implied that dithering was used AS bitrate conversion.
     It is my understanding that dithering is adding little chunks of noise to a wav which causes the already present noise to blend in better and makes it less audible.  It is also my understanding that this should be used each time after any amplitude adjustment (such as normalizing, eq, compression, or any time you are changing the "volume" of the wav or part of the wav), and again anytime you change either the bitrate or samplerate.  It is also my understanding that the only noise that this will correct is distortion which is caused by the above.
     I would like to hear from thoes who are more educated on this subject than I, and clear this up for all of us.

Thanks in advance!

Matt

[jk labs]

It's a huge topic and it's not all about noise. In a way you can say that dithering is about adding noise so as to make the digital signal better resemble the analog signal:   
say you have a 16 bit converter. The analog input is just a fraction too weak to be detected. It lives down in the 17th bit. If you now add noise to the signal the combined signal will be detected now and then. In fact you get something that resemble the analog signal. But now there is added noise here as well.

So dithering is about adding the proper amount of noise during the process of data truncation in order to get a better, more accurate representation of the signal.     

Noiseshaping is about dithering with noise of such a spectrum (and nature) that the noise added
to the recording is the least amount objectionable. 


Thing is when you change the bitrate, you are often forced to form products and excecute huge summations. This prosessing ought to be done at a much higher resolution than the original data. Say 32 bits are used. And then at the end, when all post processing at this higher resolution is done, you must truncate down to target bitdepth ....

Edited Ia: In light of what you wrote about normalization etc I should just have posted in your support on the issue of samplerate conversion: Truncating the bitdepth has nothing to do with inserting or removing new timesamples.

Edited Ib: All operations on the data involving some form of multiplication cause the bitdepth to increase. Dithering must be performed *every* time that data is truncated to a lesser bitdepth.   

PS: You can't mask noise that's already on the recording. But you can reduce the quantization noise that brute force chopping off would have caused... 
 

[MattD]

Ok... first off I would like to say I claim to be NO expert on dithering, but I thought I would open up a discussion on it.  I was looking through a few of the posts in this thread relating to dithering.  There were a few places where it was stated or implied that dithering was used AS bitrate conversion.

It is used "during" bit depth conversion. You are correct in your assessment.

It is also my understanding that this should be used each time after any amplitude adjustment (such as normalizing, eq, compression, or any time you are changing the "volume" of the wav or part of the wav), and again anytime you change either the bitrate or samplerate.

This is incorrect. Dithering should only be performed once, generally as the FINAL STEP before creating the master.

[mmmatt]

Thing is when you change the bitrate, you are often forced to form products and excecute huge summations. This prosessing ought to be done at a much higher resolution than the original data. Say 32 bits are used. And then at the end, when all post processing at this higher resolution is done, you must truncate down to target bitdepth ....

Are you suggesting that we should convert to a higher bitrate than the original recording, then perform all post processing, then convert to target bitrate just for the final product?

Edited Ib: All operations on the data involving some form of multiplication cause the bitdepth to increase. Dithering must be performed *every* time that data is truncated to a lesser bitdepth. 

 

This is incorrect. Dithering should only be performed once, generally as the FINAL STEP before creating the master.

Seems we have differing opinions here.  Anyone else want to chime in?  Thanks guys!  +T

Matt

[MattD]

This is incorrect. Dithering should only be performed once, generally as the FINAL STEP before creating the master.

Seems we have differing opinions here.  Anyone else want to chime in?  Thanks guys!  +T

I don't think I'm diplomatic enough to say this without sounding like an ass, but please don't take this as a personal attack. My statement above was not opinion. Dithering introduces quantization noise. Dithering multiple times compounds the noise.

[jk labs]

Thing is when you change the bitrate, you are often forced to form products and excecute huge summations. This prosessing ought to be done at a much higher resolution than the original data. Say 32 bits are used. And then at the end, when all post processing at this higher resolution is done, you must truncate down to target bitdepth ....

Are you suggesting that we should convert to a higher bitrate than the original recording, then perform all post processing, then convert to target bitrate just for the final product?

.....

 Thanks guys!  +T

Matt

(Who tricked me into using the word bitrate above?    It doesn't really belong to this discussion.)

My suggestion is that you leave the samplerate alone unless you have a good reason for changing it.
But if the need for a different samplerate arises then the need for truncation and dithering is present.

Every time you manipulate the data and save to file, and the process of saving involves truncating valid data, you must dither. And it's not optional.  Yes, that means a lot of dithering. This is _the_ reason for the practice that
one plans the entire mastering process well and do all mastering while keeping the intermediate data at a higher
bit resolution. That way you only have to dither once: during the final save.

Democrats may see things differently ... 

[mmmatt]

Ok... I think maybe you guys are more on the same lines than I thought.  Let me back up a bit.  I currently record to a jb3, and now that I have recently acquired decent software for samplerate conversion/dithering (wavelab essential) I am back to recording at 48/16.  My reason for popping into this part of the world (computer recording) is that I'm planning on blowing the dust off the laptop and attempting a 24bit recording next time I run my matrix rig.
     Now, when I do an ambient recording I don't typically touch my recordings.  I just change samplerate (if I recorded at 48/16) and cut the sucker up.  Sometimes I get a bass hum or something and I have to do more, but it is rare as I typically use a rolloff on the mics when I see that comming.
     With my matrix rig it is a completely different story.  I often do some global processing, and then cut, and then do processing on each file individually.  The last three things I do are convert from 48/16 to 44.1/16, dither, and save.  If I do any amplitued adjusting I dither between each step.  So in my situation I may dither the file 3 or more times.  I think it is safe to say that I am dithering too much, and as long as I complete my processing prior to saving I only need to dither prior to a save.
     If this is the case, then I will be saving myself a lot of dither time!  However I will have to rethink my global processing step because this will require an additional dither step and potentially add unwanted noise.

Am I on track now?

Matt

[rustoleum]

Process (EQ, Normalize, etc).  Then Resample.  Then Dither.

[mmmatt]

Thank you JK and Matt.  This is starting to make sence to me now.  I will happily put together a blind test if someone has a way to host it.  I could take 2 30sec chunks from a wav... one that has quiet spots and one that has more drive and do dithering your way, my (old) way, and none at all.  I have only recently started dithering anythingl, and I do notice a difference between dithering and not.  Mostly through my headphones though (sennheiser 280s) as my playback (nakamichi A/V reciever > polk s-10s) is very forgiving and anything but acurate.

Matt

[Scooter]

In Wavelab, ANY time you move a fader, insert a plugin, or other certain types of things, it is upsampling to 32 bits, then downsampling again.  So if you render multiple times w/ processing, you must dither each time.  I like the earlier suggestion of saving in 32 bit format if you think you might need to process again at a later time, good idea!  This seems like it would negate then need for multiple dithers when you process multiple times.  As far as if you can hear the effect of multiple dithers, good question .

[mmmatt]

I like the 32bit suggestion as well, but my stone age 18GB-HD computer can't handle multiple versions of files at 16bit let alone 32.  New computer soon and I won't worry about that, but for now I'm just going to start what I finish    I rarely do global processing anyway, it is ussually file by file.

   ***Side note*** this would be another good reason for TS.com to have a torrent tracker.  It would be nice to torrent tests like this!
35-40 MB on a torrent would run fast.

Matt

[jk labs]

In Wavelab, ANY time you move a fader, insert a plugin, or other certain types of things, it is upsampling to 32 bits, then downsampling again.  So if you render multiple times w/ processing, you must dither each time.  I like the earlier suggestion of saving in 32 bit format if you think you might need to process again at a later time, good idea!  This seems like it would negate then need for multiple dithers when you process multiple times.  As far as if you can hear the effect of multiple dithers, good question .

We'll see We know that dither lessens the severity of the quantization noise that occurs when truncating from analog or high resolution, to lower resolution.  The data that is of interest to us was for sure truncated and dithered in the ADC. So the final dithering will be added to the initial dithering irrespective of how much we try to resist it.

Now comes the fun part. The different noiseshapes used when dithering might be "incompatible".. That is when you dither on top of an already dithered signal bad things _could_ happen.

This is something to be aware of if dithering is applied several times using fancy noiseshaping. "Dithered 4 times" might be very different from "dithered four times".

To be of value any test must keep track of the noiseshaping algortihm(s) used.

[mmmatt]

Well... my std wavelab "internal" dithering doesn't give anything to scientiffic... just noise 1 or 2 and shaping 1,2 , or 3.  I could certainly do the same and then do mix-n-match as a 3rd test.  The other thing we could do to get this around is just e-mail the files to a phony yahoo e-mail account, then just share the password and login.  I think the yahoo acounts are 100 mb now, and I'm sure they would handle a 5mb file.

Matt

[mmmatt]

ok... the place to get the files is through a gmail act.  www.gmail.com
login is dither.test
password is tsdotcom
    I have diter test "a" up and b" in just a few minutes.  They can be downloaded from the "sent" folder.


Ok... here is the test!
2 files "A" is a rock jam and "B" is a quiet acoustic jam.  "A" is
from a matrix that had a fair amount of noise to start with coming off
the soundboard feed.  "B" is mics only and is as clean as I have
(sorry no u87's here!)

"A"  akg391's (on stage corners facing audience) + SBD > Mackie ONYX
1220 > ua-5 (for A/D only) > jb3@48/16

"B" akg391's DIN, DFC, 15' > ua-5 (d-mod only) > jb3@48/16

processes are as follows (please note I wasn't trying to improve the
sound... just random processes!)
all files:
fade in
fade out
normalize to -0.1
compress @ 2 to 1 w/-20db threshold
normalize to -0.4
master volume to +.15
convert samplerate to 44.1/16

In each of the 2 file groups I produced 4 seperate wavs
one of each has the folowing

no dithering at all

dither with the same noise/shaping after every change except the fades
(5 rounds of dither)

dither with different noise/shaping after every change except fades (5
different combinations of dither)

dither only at the end of the process during the save sequence

Bonus points is you can gues the bands, and the songs but at least song "a" is a nobrainer.

Good luck!!!
Matt


Matt

[mmmatt]

they are all up now

Matt

[dklein]

good work man.  I'll give these a go.  I'd recommend using this tool for your comparisons if you're playing back on a pc/laptop source
http://www.kikeg.arrakis.es/winabx/

Here's some other sources in you're not a windows user http://www.pcabx.com/

I must admit, I'm a little bit skeptical about this practice of dithering after each step as most software already bumps up the resolution for processing and dithers back down automatically.  But let's give it a listen first!

[mmmatt]

As were listening there is one more thing to consider.  Dithering (here again I'm no authority) in my understanding not only uses it little bits of noise to soften the noise created by some forms of processing, but it also brings subtle nuances out in the source that would otherwise be lost.  By adding a chunk of noise to the ends of these nuances, they become audible.
     Ok JT, or Matt (or anyone else who knows this stuff)... now say it right for me but I think I'm close!  Point being that as we are listening for less noise, we should also listen for more music.  Different versions of the test may have strengths in one end but not the other or there may be one best way to achieve both benefits of dithering.  I guess we will find out!!!
     I think I'm going to post a link to this on another thread.  I think many non-lappy tapers (like me) don't often make it to this side of the sight.

Matt

[jk labs]

As were listening there is one more thing to consider.  Dithering (here again I'm no authority) in my understanding not only uses it little bits of noise to soften the noise created by some forms of processing, but it also brings subtle nuances out in the source that would otherwise be lost.  By adding a chunk of noise to the ends of these nuances, they become audible.
     Ok JT, or Matt (or anyone else who knows this stuff)... now say it right for me but I think I'm close!  Point being that as we are listening for less noise, we should also listen for more music.  Different versions of the test may have strengths in one end but not the other or there may be one best way to achieve both benefits of dithering.  I guess we will find out!!!
     I think I'm going to post a link to this on another thread.  I think many non-lappy tapers (like me) don't often make it to this side of the sight.

Matt

Seems I'll need a little "dither strategy" paper anyway so this is my take on this:


The MS Excel graph below attempts to show an analog sine wave (in green). The maximum is at "17", thinking this is suitable to demo the performance of an ADC with 32 discrete levels (5 bit converter). The frequency of the analog signal is about 1% of the sample frequency.  The sine wave contains no noise.

This analog signal is input to an ADC implemented in MS Excel spreadsheet.

The ADC in question is assumed to be perfect: The ADC has a totally stable voltage reference against which it measures the signal. The ADC does not inject RF into the analog part. The ADC does not mess with the power supply for the analog stage. The ADC has a totally isolated digital ground. The clock has no jitter. The ADC does not add any jitter to the clock signal. There is no leakage of the clock signal into the analog signal. The analog signal does not modulate the clock. The sample and hold holds a perfect representation of the analog amplitude until this amplitude is sampled. The ADC does not deviate from linear in its conversion of small or large signals.  The ADC input gate adds no noise to the signal, there is no capacitive modulation,... ad infinitum....   

The result of this perfect ADC, processing the green analog signal, is represented in the red dataset. The output is a sequential series of integers in the range -15 to +16 sampled at a fixed interval. The sample is valid for that one instant in time.

We see that there is a difference in the amplitude of the two representations. Truncating the data causes this discrepancy. This is the quantization error. Eyeballing the graph we see that it is less than on bit in amplitude. In fact, it's for all practical purposes triangular and so eats up 50% of the least bit.
Ooops, we just lost 3 dB dynamic range right there.   

But we also see that the quantization error itself stems from changes in the analog data. That is, the quantization error is *correlated* with the input signal. Our ADC, when processing analog data, introduces it. That's bad. Not only do we loose 3 dB dynamic range but we find scads of new harmonics in our digital representation of the original analog signal. So our perfect ADC alters the very signal!

What is worth noting is that the exact same thing happens if you do *any* floating point processing on already digitized data and then proceed to store the result as an integer in any form. The severity of the effect depends on the bitdepth and scales with the relative size of the least significant bit.        

Dithering means to add noise to the analog signal. The quantization error will now correlate with the signal + the dithering noise! Carefully selecting the dither amplitude to +- 1/2 bit (can be "explained" with basis in the graph) derails the quantization error quite effectively. The error is now correlated with the noise! I.e. correct dithering manages to break the correlation between the desired signal and the quantization noise. Our original signal is now better represented in the digital domain but the drawback is that more noise is present in the digitized dataset.

For dithering to work, the noise must not be correlated between samples (it must not be "signal like"). This means it must contain high frequency components causing a different noise value at every sample.
   
White noise is something that the ear and brain is used to. It's also harmless on subsequent processing and it is easy to generate in both the analog and digital domain. So white noise is the basic form of noise used with dithering.

A strategy for reducing the perceived level of noise is to put it in a region of the spectrum where the ear is less sensitive to acoustic energy. By sending all the noise energy up in frequency one attempts to bring the quantization error out of the audible range. This is the idea behind noiseshaping.

PS. The MS Excel graph contains some oddities. The sine wave is clearly distorted and the vertical bars are
offset by half a time step to the right. The first problem is odd in that all data is generated by an Excel formula. The latter is corrected by some option setting I'll find one day...   

Edited: reduced the size of the image to suitable size.

[dklein]

Well I guess I'll be first to go on the record and say I'm not hearing differences.  While I *thought* I heard some, I could not reliably distinguish using that ABX software I linked to above (did ok with samples b1 and b4)..

Why?  Some possible explanations.

- there is already so much noise in both recordings that the incremental noise from dithering is still below the ambient noise floor
- there are no 'subtle sounds' being brought in from outside of the 16 bit (96dB) range that are actually audible over the ambient noise
- there was so much destructive processing that the dither effect was minor compared to the other stuff
- there was no sound that decayed to nil, where the quantizing effect of no dither would be heard
- there is no undithered sample (because wavelab automatically calculates at a higher resolution and brings it back down if you're working in 16 bit mode (btw, what did you set the wavelab temp file resolution at? - check preferences, file)
- my ears are not what they used to be (I'm 37)

Playback was laptop > emagic A62 (digital out) > Lucid DA9624 > Linn/Naim hifi (should be good enough)

I think that we'd need studio quality material and some real decaying sounds to bring out the differences.  Which kinda says that for our purposes this may not be a big deal.  When I say 'for our purposes' I mean field recording in less than ideal circumstances where we have very little conrol over the environment and there's usually tons of background noise.

Who's hearing differences?  And if you are, I challenge you to do it again with the ABX software and hide the results.  Run 5-10 trials and let the software tell you if you're guessing or not.  fyi - set it to ABX, load the 2 files.  All you have to do is say whether X (randomly selected by the software) is the same as A or B.  Then click 'next trial' after you've decided.  Check the 'hide results' button - it'll keep you honest.

[mmmatt]

- there is no undithered sample (because wavelab automatically calculates at a higher resolution and brings it back down if you're working in 16 bit mode (btw, what did you set the wavelab temp file resolution at? - check preferences, file)

32 bit float temp file.  I've never adjusted this so it must be the default.

Matt

[BWolf]

just like to say thanks for that information.  i'm an electrical engineering student and that was extremely enlightening.  thanks

taken so many systems and signals/computer processing and there isn't one that is tied to music/audio in any way, so i always had a hard time picturing the digital audio world.  but that put a lot into perspective.  thanks.

[mmmatt]

anybody hear anything?  I will post the key on monday so get the samples if you are interested.

Matt

[mmmatt]

Well... I think this test was a bust.  However here is the key.

Clip "A"  Fire on the Mountain   The Big Wu   wu2004-10-15 

1  different dithers at each step
2 No dithering whatsoever
3 dither same at each step
4 dither at the end

Clip "B"  Rosalie McFall  The Clumsy Lovers  TCL2004-10-14

1 dither at the end
2 dither same at each step
3 No dithering whatsoever
4 different dithers at each step

Naturally I wasn't as blind as the rest, but with my blinders on I heard subtle differences between a few.  Not so much in the noise factor, but in the overall presence of the music that was there.  I felt my original method of dithering between each step was the best.  I heard it most on the glass clink at the begining of clip B.  Who knows though... maybe I just wanted to hear something.  Either way it was so subtle that one would never notice through typical listening applications.

Matt