July 22, 2021
 by Paul McGowan

The dictionary has several meanings to the word dither. The most common is to be indecisive. The second meaning is more related to what interests us: to add white noise.

Dither is the addition of randomness into low-level signals. Surprisingly, we do this to lower distortion of predictable errors.

Here’s a great example from a Photoshop help book on digital photography.

The first image hasn’t enough low-level information to smoothly transition between the differing shades in the bird’s wings. By adding some random noise—dither—we can smooth out the image as in the second example.

In digital audio, the problems of quantization (converting a steady-state analog signal into a series of fixed values) are essentially the same: a predictable error between steps generates measurable distortion. If we randomize the errors by adding in a bit of dither, then a few things happen for the better. First, our chances of hitting the right transition levels are bettered, and second, our ear is forgiving of random and uncorrelated events.

We can easily ignore a touch of noise but it is much more difficult to ignore distortion. Look again at the two pictures. I bet you didn’t even notice the 2nd picture is slightly noisier than the first.

Nearly every CD you have ever listened to has some amount of dither added to it and for the better.

And all along you thought we go out of our way to lower noise every chance we get.

If you want to dig a bit deeper into the how and why we add noise to lower distortion, there’s a great Wikipedia article on it you can access here.

Subscribe to Paul's Posts

22 comments on “Dither”

  1. The last time I remember quantisation errors in the form of banding on a photo was with a Canon 350D that had an 8mp sensor.

    The question is, if for example Hyperion who master in 24/96 PCM and Linn in 24/192 PCM, and I buy files in those bitrates, presumably there is no quantisation involved and hence no dither added? Could life possibly be that simple? Surely not?

    1. Good questions, Distortions are in every digital signal. Today’s microphone recorders are direct to DSD. I have played around a bit with the AIX recordings and really could not hear much of a difference between their PCM recordings and DSD using full multi-channel setup. To me the PCM recordings where a tiny bit softer than the DSD. This whole DSD thing is a very complex field of electronics to comprehend. I have to this day wondered why SACD volume is so much lower that standard CD volume are they hiding something?
      I’m a retired electronics tech and the chip manufactures specs and technical application books makes my head spin when trying to get a handle on what is really happening!

      1. Good morning David!
        To put it to you cempally, PCM has to lower the gain so there isn’t any digital noise to be heard way up there.
        DSD, uses a scheme that’s called PDM or post dither modulation.
        But a DSD file, contains way more sonic information then a PCM or post code modulation file does.
        And yes, you’re right.
        The technical side of DSD, is a little more complicated.
        It’s kind of hard for me to follow it as well.
        I do have some understanding of it, but not a very deep understanding of it, like I wish I did.
        I’m a 49 year old blind Audiophile that loves both music and vacuum tubes.

        1. Hi John,
          Ah Ha, you rattled my brain a bit and now remember reading a bit about PCM and the lower gain. At the time I never put 2 and 2 together. Thanks, will see if I can dig out that article I was reading, wow that was a long time ago..

  2. As long as it sounds good, I’m fine with the measures on the way.

    This story might just be puzzling for those who thought that digital means the accurate and unaltered reproduction of the input.

    1. “But things ain’t quite that simple.” — Pete Townsend

      I seem to be on a bit of a The Who jag. Things could be worse Maximum R&B!

  3. Our brain is capable of detecting 3 picosecond difference, sound direction. Nanosecond accuracy is about the best commercial products can do. I’m with you jazznut we have what is we have and it is good enough for current day and age. Would think the limiting factor on what we all hear is in the speakers we use.

    1. David you are right about human detection of sound arrival times, however, the are quite a few DAC’s out there that have 20 picoseconds or less of jitter. This, of course, has nothing to due with dither which is all about quantization error.

  4. With the photo’s, the error in the one that is not dithered is visible at the shown magnification. The noise in the dithered picture only becomes visible if you enlarge the picture. Would this be similar for music: digital errors are easy to hear at low volume and dithering at high volume?

    1. I think I had that with the Auralic Aries Mini, which I was using in my office directly into a power amplifier (i.e. using the internal DAC and volume control). The internal DAC has a very high voltage output and using the internal volume control I had to have it set at around 10%. It sounded terrible. I assume it was a quantisation issue with the volume control. I fixed it with a pair of Rothwell attenuators to drop the output voltage, and Auralic fixed it with a software update, which I presume was some sort of dithering.

  5. Failing to dither digital audio truncations is like not using AC bias on a tape recorder. It should never be optional and ought to simply be an integral part of engineering design.

  6. I’m your tricksy Uncle Skinner.
    Now let’s make this thing a winner!
    As I dither about, dither about, dither about.

    The Recording Engs did mix things,
    Now the Master’s here to fix things!
    Dithring about, dithring about, dither about.

    Tweak up the balance.
    Toss out compressers.
    Dither about, dither about, dither about.

    With apologies to the Ox. Oh, wait, he’s dead.
    [Emily Litella font on] “Never mind.” [Emily Litella font off]

  7. In my opinion dither and quantization error ( QE ) play a much bigger roll in video digitization than they do in digital audio. The effects of quantization limits on the color pallet are very obvious in the image on the left, however, it is also very obvious to me ( and I wear progressive lenses in my eye glasses ) the even with the small image size on my laptop I can tell that the image on the left is sharper than the dither one the right. Also another method can also be applied to the digital date to reduce the effect of QE errors and that is noise shaping ( https://en.wikipedia.org/wiki/Noise_shaping ).

    In my opinion in digital audio QE is not much of a problem compared to the effects of low sampling rates and jitter.

    1. The truism goes that there there is virtue is seeing (and hearing) things as they are, warts and all. But when the warts are artifacts of of the digital recording process, optical and/or sonic, of our own making through still imperfect technologies, is doing our best to correct those warts a bad thing? It is just my opinion, but I am willing to exchange a little sharpness for continuity. That is how I see and hear (imperfectly) with my neurology, which in all likelihood is slightly different from yours. I’m not a theoretical physicist, but from my limited understanding of such things, baring a breakthrough in the ‘Theory Of Everything’, above the quantum level the macroscopic world outside of computers and such behaves more like analog than digital. Digital recording and playback has come a long way, but like analog (before and still), it is a work in progress.

  8. The effort has always been to reduce all types of distortion in the analogue realm. In digital audio it is the reverse. Distortion is added to negate another distortion. Runs contrary to all logic and reason. Obviously digital is a significantly compromised system and needs a complete overhaul. No wonder in it’s present state only those who can not hear it’s distortions like it. Two evils do not make a good. Regards.

    1. Oliver, If only it was that simple. I’m sure there mathematicians and electrical engineers all over the world that would love to hear how you propose overhauling digital audio. 😉 . Digitization of any kind of signal that varies in amplitude and frequency with respect to time is a mathematical process that must adhere to certain mathematical boundary conditions. Current digitization processes adhere to the Shannon / Nyquist sampling theorem. To go beyond that into non Shannon / Nyquist sampling is extremely difficult and I personally do not know of any practical applications that can do this. MQA tried to claim that they could to this, but that has not been widely accepted.

      The biggest mistake that in current digital audio is that the first practical consumer application of digitization ( the CD ) made a very stupid mistake by not understanding the importance of frequencies above 20 kHz in what we hear and thus chosing a sampling rate that was too low for good audio quality. That legacy still hinders digital audio today more than three decades after its launch. If you think you know how to restart digital audio go for it!

  9. Yup. This is mostly about our brain’s tolerance for different levels of distortion and the range of it.
    As most of us get older we supposedly tolerate higher frequencies and brighter recordings because as we age we naturally lose depth in our mid range hearing. It is more commonly known as the cocktail Party affect. One may notice that when in a loud environment it becomes harder to hear the person next to you trying to talk to you. 😉

Leave a Reply

Stop by for a tour:
Mon-Fri, 8:30am-5pm MST

4865 Sterling Dr.
Boulder, CO 80301

Join the hi-fi family

Stop by for a tour:
4865 Sterling Dr.
Boulder, CO 80301

Join the hi-fi family

linkedin facebook pinterest youtube rss twitter instagram facebook-blank rss-blank linkedin-blank pinterest youtube twitter instagram