Can someone explain to me what dithering is?

[Drtro]

I'm sure you can google and the first result will be a definition or Wikipedia article. But I'm too lazy to copy/paste.

[johnnyv]

Completely wrong. But I like above answer best because if I try to explain it I'll get it wrong and it will start a 3 page thread.

[ShadowAMD]

Originally Posted by Billly

From what i understand it is the regaining of lost information when an analog signal is converted into digital or back in to digital? Or is that completely wrong? I really have no idea.

It adds noise due to issues in quantisation when sampling.

[Lotus 7]

Originally Posted by Billly

From what i understand it is the regaining of lost information when an analog signal is converted into digital or back in to digital? Or is that completely wrong? I really have no idea.

Not even close, but don't worry about it.

The very short answer: It's a method to convert larger digital sample sizes to smaller sizes, while trying to preserve some of the detail carried in the bits that are removed. There are various methods that are used, most involving adding a very low level, shaped random noise spectrum.

The Wikipedia listing is detailed but is also very "dry" and includes excruciating detail that will turn off almost everyone.

For a shorter, more understandable explanation see: What is dither?

[Billly]

thanks so much guys this really helped a lot

[audioexmachina]

Originally Posted by Billly

thanks so much guys this really helped a lot

Short answer (my try):

The act of reducing bit depth introduces background noise that is modulated by the original waveform, resulting unpleasant. Dither modifies the reduction process so that the generated noise is not modulated by the source, resulting less annoying.

Noise-shaping changes that noise's EQ so that more energy is moved to frequencies that are less audible, resulting in an apparent quieter noisefloor.

[SochiTag]

Try here: iZotope Insider Tips: Introduction To Dither - YouTube

[joe_04_04]

I could never wrap my head around that technical jargon, just know that it is something that is necessary when bouncing your 24 bit mixes to 16. Its the most boring plug that will ever be in your signal chain, unless you are recording to disk in which case pro tools (if you use it) automatically adds dither upon exporting the bounced region (clip in pt 10).

[ShadowAMD]

Originally Posted by audioexmachina

Short answer (my try):

The act of reducing bit depth introduces background noise that is modulated by the original waveform, resulting unpleasant. Dither modifies the reduction process so that the generated noise is not modulated by the source, resulting less annoying.

Noise-shaping changes that noise's EQ so that more energy is moved to frequencies that are less audible, resulting in an apparent quieter noisefloor.

Sorry but your wrong, I can pull you towards 50 whitepaper's saying that.

[audioexmachina]

Originally Posted by ShadowAMD

Sorry but your wrong, I can pull you towards 50 whitepaper's saying that.

I'd be interested in knowing what these papers say: would you please post a quick summary in a few words, if you have time?

I can assure you, having worked on the subject for a long time, that the purpose of dither, in engineering, is the decorrelation between the quantization noise and the original signal.

[ShadowAMD]

Originally Posted by audioexmachina

I'd be interested in knowing what these papers say: would you please post a quick summary in a few words, if you have time?

I can assure you, having worked on the subject for a long time, that the purpose of dither, in engineering, is the decorrelation between the quantization noise and the original signal.

I've done audio R&D for over 10 years, so nice to meet other guys..

"Dither is an intentionally applied form of noise used to randomize quantization error, preventing large-scale patterns such as color banding in images. Dither is routinely used in processing of both digital audio and digital video data, and is often one of the last stages of audio production to compact disc."

Dither - Wikipedia, the free encyclopedia

https://docs.google.com/viewer?a=v&q...7dPCoUvQ&pli=1

Here is just a few, you only have to google it and they are spot on!

[perform/record]

Originally Posted by Lotus 7

...

For a shorter, more understandable explanation see: What is dither?

EDIT:

Thanks, this was interesting.

[audioexmachina]

Originally Posted by ShadowAMD

I've done audio R&D for over 10 years, so nice to meet other guys..

My pleasure as well! Looks like I'm older so won't expose my mileage in the field :-) I think the problem is just that a rather clear mathematical topic is referenced by many sources in natural language instead of formulas, leading to possible misunderstandings.

"Dither is an intentionally applied form of noise used to randomize quantization error, preventing large-scale patterns such as color banding in images. Dither is routinely used in processing of both digital audio and digital video data, and is often one of the last stages of audio production to compact disc."

Right, "to randomize quantization error" references decorrelating the quantization error and the original signal.

This book (just the first coming out from google, never seen before), page 81 ( 2.8.10 Dither for Digital Interface Handbook, Third EditionAuthor(by John Watkinson and Francis Rumsey) ):

"If the quantizing error can be decorrelated from the input in some way, the system can remain linear. Dither performs the job of decorrelation"

Dither - Wikipedia, the free encyclopedia

Wiki says the same:

"If that error is repeating and correlated to the signal, the error that results is repeating, cyclical, and mathematically determinable. In some fields, especially where the receptor is sensitive to such artifacts, cyclical errors yield undesirable artifacts. In these fields dither results in less determinable artifacts."

That's obtaining decorrelation by whitening the quantization noise.

https://docs.google.com/viewer?a=v&q...7dPCoUvQ&pli=1

Same here too:
"When we now remove the last 8 bits to finally turn this signal into a 16 bit signal you'll see that we no longer have determineable "stair steps" as we did when just truncating, but we rather have random quantization fluctuations of single bits here and there."

Also, this last paper has a flaw: it suggest that dither noise should peak at the LSB level of the output signal: "In our working example we're dithering 24bits to 16 bits, so we're going to add 9 bits of noise."

That's a misundertanding of the function of dithering: that way one could dither a 16bit signal starting with a 16bit signal and random-flipping the LSB :-)

A proper example, IMHO, would use sub-LSB amplitude, such as 0.5bit peak levels for dither noise, showing that adding noise BEFORE the quantizer (even below the minimum sensitivity level of the quantizer itself) is the key to dithering. We know it's okay to peak dither up to the LSB, depending on the statistical distribution shape, but the paper suggest that as a REQUIRED condition. Which is not true.

[ShadowAMD]

Originally Posted by audioexmachina

My pleasure as well! Looks like I'm older so won't expose my mileage in the field :-) I think the problem is just that a rather clear mathematical topic is referenced by many sources in natural language instead of formulas, leading to possible misunderstandings.




Right, "to randomize quantization error" references decorrelating the quantization error and the original signal.

This book (just the first coming out from google, never seen before), page 81 ( 2.8.10 Dither for Digital Interface Handbook, Third EditionAuthor(by John Watkinson and Francis Rumsey) ):

"If the quantizing error can be decorrelated from the input in some way, the system can remain linear. Dither performs the job of decorrelation"





Wiki says the same:

"If that error is repeating and correlated to the signal, the error that results is repeating, cyclical, and mathematically determinable. In some fields, especially where the receptor is sensitive to such artifacts, cyclical errors yield undesirable artifacts. In these fields dither results in less determinable artifacts."

That's obtaining decorrelation by whitening the quantization noise.




Same here too:
"When we now remove the last 8 bits to finally turn this signal into a 16 bit signal you'll see that we no longer have determineable "stair steps" as we did when just truncating, but we rather have random quantization fluctuations of single bits here and there."

Also, this last paper has a flaw: it suggest that dither noise should peak at the LSB level of the output signal: "In our working example we're dithering 24bits to 16 bits, so we're going to add 9 bits of noise."

That's a misundertanding of the function of dithering: that way one could dither a 16bit signal starting with a 16bit signal and random-flipping the LSB :-)

A proper example, IMHO, would use sub-LSB amplitude, such as 0.5bit peak levels for dither noise, showing that adding noise BEFORE the quantizer (even below the minimum sensitivity level of the quantizer itself) is the key to dithering. We know it's okay to peak dither up to the LSB, depending on the statistical distribution shape, but the paper suggest that as a REQUIRED condition. Which is not true.

Umm, but I agree with the above LOL! I guess I must of just took the wrong end of the meaning in your post. It's hard not to be pedantic on the subject sorry :D

[Bob Olhsson]

My recent tests have shown that full level flat dither actually is required but if the signal has ever been truncated without dither, sub-LSB dither levels or noise shaping can sound better. It just won't sound as good as audio that was properly dithered with full level flat dither at every truncation step.

I realize this is a bit counterintuitive. The reduced level and noise shaping were all developed at a time when most converters had truncated oversampling and most recordings were made with 18 bit converters truncated to 16 bits. Modern converter chips are really a whole new ballgame that demonstrate the theory actually was correct.

[audioexmachina]

Originally Posted by ShadowAMD

Umm, but I agree with the above LOL! I guess I must of just took the wrong end of the meaning in your post. It's hard not to be pedantic on the subject sorry :D

Yes, I'm sure we were saying the same thing :-) It doesn't hurt as that makes this thread a bit more informative (hopefully).

[audioexmachina]

Originally Posted by Bob Olhsson

My recent tests have shown that full level flat dither actually is required but if the signal has ever been truncated without dither, sub-LSB dither levels or noise shaping can sound better. It just won't sound as good as audio that was properly dithered with full level flat dither at every truncation step.

Bob, do you have any reading for the dither peak value (and shape: RPDF,TPDF,...) at "full level", during those tests?

I'm asking as the 0.5bit I cited is the theorical minimum required ONLY for the rectangular shape (RPDF). If using TPDF you'd need to peak at least at 1.0bit, and even more with other custom shapes, however, even with peaks that high, most of the time the dither signal will be at sub-LSB level, due to the distributions not being rectangular. Basically, the higher the peaks, the less they happen.

[Bob Olhsson]

Triangular 2.0 bits.

[JamesClark1991]

So lets say I record a vocalist at 24bit 48khz and then later on convert it to 16bit 44.1 should I dither it? and if I do will there be any artifacts? would it be better to just record straight to 16bit 44.1khz in the first place?

Any help appreciated

[Bob Olhsson]

If you do any digital signal processing, even a tiny gain change, you are creating 56 to 80 bit audio that needs to be dithered back to the final bit depth. I won't record anything at 16 bits for this reason.

I think a lot of plug-ins sound better at 48k and better yet at 96k so I record at those sample rates. At this point most of what people get paid for is video which would need to be converted to 48k. 48k MP3s sound better too.

The 44.1 conversion also creates 80 bit audio that needs to be dithered to 16.

[ShadowAMD]

Originally Posted by JamesClark1991

So lets say I record a vocalist at 24bit 48khz and then later on convert it to 16bit 44.1 should I dither it? and if I do will there be any artifacts? would it be better to just record straight to 16bit 44.1khz in the first place?

Any help appreciated

Record in 44.1 /24 there are many reasons, especially when tracking.

[audioexmachina]

Originally Posted by JamesClark1991

So lets say I record a vocalist at 24bit 48khz and then later on convert it to 16bit 44.1 should I dither it?

Yes, you should dither at the moment you reduce bit depth. Not later (as extra infomation from lower bits has already been lost), not earlier (as possible further processing stages may damage dither).

and if I do will there be any artifacts?

Yes, but there would be artifacts even if you don't dither. The purpose of dithering is replacing unpleasant artifacts with more acceptable artifacts. Specifically, you reduce distortion at the expense of your S/N ratio (you get a little bit of extra background noise).

would it be better to just record straight to 16bit 44.1khz in the first place?

(I'm skipping the 48<->44.1 conversion intentionally). It wouldn't be better to record at 16bit in first place. Look at it this way: that would be a bit reduction as well, from analog to 16, and you'd need to dither the input signal before it gets captured to obtain a quality comparable to a dithered 16bit coming from a 24bit source. Even doing that, you'd end up with a recording that, if processed, may loose its dithering components (see conditions at the start of this message).

Starting with hi-resolution sources and applying bit reduction only at the very last step is the correct way to preserve signal quality. Sometimes this is not possible (e.g. you're on a 32FP buss and need a roundtrip through the converters for a send/return connection to outboard gear)

I'd like to remark that all this is quite an academic perspective as in practice operating in a real world with rear devices (can I trust my hardware? and expecially my software?) introduces many other variables. Still IMHO it's not bad to keep theory as a reference.

[ShadowAMD]

Originally Posted by audioexmachina

Yes, you should dither at the moment you reduce bit depth. Not later (as extra infomation from lower bits has already been lost), not earlier (as possible further processing stages may damage dither).



Yes, but there would be artifacts even if you don't dither. The purpose of dithering is replacing unpleasant artifacts with more acceptable artifacts. Specifically, you reduce distortion at the expense of your S/N ratio (you get a little bit of extra background noise).



(I'm skipping the 48<->44.1 conversion intentionally). It wouldn't be better to record at 16bit in first place. Look at it this way: that would be a bit reduction as well, from analog to 16, and you'd need to dither the input signal before it gets captured to obtain a quality comparable to a dithered 16bit coming from a 24bit source. Even doing that, you'd end up with a recording that, if processed, may loose its dithering components (see conditions at the start of this message).

Starting with hi-resolution sources and applying bit reduction only at the very last step is the correct way to preserve signal quality. Sometimes this is not possible (e.g. you're on a 32FP buss and need a roundtrip through the converters for a send/return connection to outboard gear)

I'd like to remark that all this is quite an academic perspective as in practice operating in a real world with rear devices (can I trust my hardware? and expecially my software?) introduces many other variables. Still IMHO it's not bad to keep theory as a reference.

What this guy said !! You type it, I'll take credit ;P

[Bob Olhsson]

What kills me is that this is exactly the kind of thing computers are really good at keeping track of. It is programming incompetence and malpractice that users ever need to even think about dither. It is as dumb as an analog machine that defaulted to the bias being disabled.

[audioexmachina]

Originally Posted by Bob Olhsson

What kills me is that this is exactly the kind of thing computers are really good at keeping track of. It is programming incompetence and malpractice that users ever need to even think about dither. It is as dumb as an analog machine that defaulted to the bias being disabled.

I couldn't agree more.. That's what I see happen when products are completely designed by technicians, with real users having no role in the process.

IMHO a great product comes from users: engineers just need to figure out how to build it and make it work, then leave.