|
e what it is, Dithering
|
View this Thread in Original format
| tehlord |
Tell me what it is, if you use it and why! :D
My projects are in 16/44 btw, although perhaps some of the samples are 24bit and I don't know!
If that makes a difference. |
|
|
| Coyke |
Hi TA,
dithering is necessary for avoiding quantization errors.
I do produce in Cubase which works in 32bit float. To avoid any quantization errors while doing any bit reduction (32bit|24bit to 16bit) you should use a dithering tool, like the UV22 in my case. Getting on cd | mp3 wont let you use more than 16 bit | 44.1 Khz so its need to be processed to avoid any distortion and noise.
I cant really explain what a dither does in detail, but I'm sure someone else can. |
|
|
| Kysora |
The way I understand it, when samples are changed to a lower bit rate, some of the data in the original sample has to be modified or removed, which causes distortion. Dithering randomizes the data that's changed, which causes noise instead of distortion, which is much less noticeable.
That's a really vague description of what it does, I'm not entirely positive of how to correctly apply it to your tracks. |
|
|
| Derivative |
You don't have to worry about dither.
If you want to know what dither is you first have to know what quantisation is so you need to know what sampling is, how a digital 'signal' is different from an analogue signal and how thats different from a sound wave which is a travelling wave in air consisting of cyclical variations in air pressure.
Starting Principles:
So start with the sound wave. When you record an instrument like a guitar or your own voice you sing into a microphone transducer which is like a thin membrane (diaphragm) that vibrates in sympathy with the sound waves you create. The membrane is attached to a magnet which moves with the diaphragm in and out of a coil. This is a simple electromagnet and this type of microphone is called a dynamic microphone. As the magnet moves in and out it induces a varying current through electromagnetic induction and this electrical signal is analogous to the sound wave because its amplitude and frequency is the same as the sound wave. This is called an analog signal.
When we sample an analogue signal, it passes through a digital to analogue converter which takes 'snapshots' of the amplitude of the signal a fixed number of times per second. In a 44.1khz samplerate converter this will be 44,100 times per second. The ratio of amplitude is determined by the bit depth. This process of 'digitizing' an analog signal is called 'quantization'. Note that you can create a signal entirely digitally with no analogue source (like softsynths). Its just math and is an abstraction.
Once converted, you have reconstructed the analogue signal as a series of fixed points. To convert it back into a medium which can drive a speaker transducer you have to put this information into a digital to analogue converter which literally 'join the dots' and turns it back into a varying electrical (analog) signal. This is done by creating a series of pulses which coincide with each sample and then passes this sequence of pulses through a reconstruction filter using some type of high order polynomial interpolation to join the pulses smoothly.
This analog signal then goes down a cable to an amplifier which uses mains electricity to make the low amplitude analog signal into a high amplitude analog signal. This then goes to a speaker which works like a microphone in reverse. Its a moving coil electromagnet connected to a 'membrane' (the cone) and the analog signal drives the cone making it vibrate with the amplitude and frequency of the signal. This sets up a longitudinal wave with analagous variation in air pressure which makes your ear drum resonate and which your brain perceives as sound.
Dither comes in during the part of this process concerning digital signals only.
Quantization Error:
So take a 24 bit/44.1 khz waveform and imagine it as a series of dots joined together so that it looks like an infinitely continuous wave. The vertical space between dots is determined by bitdepth. The horizontal space between dots is determined by samplerate. The higher the bit depth and samplerate, the smaller the space between dots.
Now if you throw away the lowest 8 bits to make a 16 bit/44.1khz recording the dots will no longer be in the same place they will be spaced further apart vertically. This difference in sample position is called a Quantization Error. Its a mathematical error that generates a very low level noise that is correlated to the signal. Use a bitcrusher on any sound and keep throwing away bits so you can hear what this error sounds like.
Dither:
Dither is a randomly generated, stereo noise source that is completely uncorelated to the signal. You use this noise to decorelate the mathematical error described above. Quantization noise in a 24 > 16 bit recording manifests at very very low level. Its down -90 to -120dB so its almost entirely inaudible except under certain listening conditions.
Its easier to understand dither in terms of images so I suggest looking at the wiki page on dither which has the picture of the cat. The first picture is a 24 bit bitmap or something so it has a colour palette of 16 million discrete levels of colour. The second image is what happens when you throw away 16 bits and have an 8 bit bmp which has a colour palette of 256 discrete levels of colour. Alot of detail is lost and you start to see area of the picture no longer graduate but step to the next colour.
The final image adds a random noise source to the image before throwing away 16 bits worth of colour information. Note how much detail is retained even though it has the same amount of colours as the 8 bit bitmap.
Now the difference between 24 bit and 16 bit is not big. Its actually tiny so the effect dither has on digital audio going from 24 bit to 16 bit is also very small. Use the wiki page to understand the principle if you want to but it doesn't make a dramatic difference. Actually, you probably won't hear much of a difference and theres a million things you should be seeking to improve before you consider dithering. Learning how to perform, mix and record properly should be the top 3 concerns of most musicians. Leave the dither stuff to the engineers.
UV22 (as mentioned above) is a noise shaping algorithm. Noise shaping dither is like a whole discipline unto itself and is based on the idea of shifting dither noise into audio bands where it is less perceptible in the programme material and otherwise generate only enough dither to decorelate the error and not more than is necessary. But its already borderline inaudible anyway under normal listening conditions so don't worry about it. I don't even bother using dither because I don't understand enough about noise shaping or quantization error to properly understand what audio bands I want to shift dither noise into or away from. I don't use instruments sensitive enough to detect the presence and magnitude of quantization noise and couldn't recognise it in a complex audio signal anyway. By extension, I don't know much dither is too much or not enough or even when its working or not. I'd be flying blind and I hate doing stuff without knowing how its working so I don't bother. I don't want to be an engineer. I want to be a musician. I assume its the same with you folks too. |
|
|
| tehlord |
Great answer Derivative. Luckily I already understood bitrates and sample depths so after your explanation it's pretty straightforward.
If my projects are all created in 16/44 there's nothing I need to consider anyway I guess.
I'm pretty happy with the way my mixes are coming out these days but I've been wondering about the benefits of rendering a final stereo mix to 32 bit float and 'mastering' (let's call it louding or awesomizing) down to 16 bit again. Obviously any samples that are alreaady 16/44 aren't going to benefit, but do my VST channels benefit from being rendered at a higher bitrate, or should my entire project be set to say 32/44 to start with? |
|
|
| Morvan |
| On GearSlutz they boast themselves with being able to distinguish the Voxengo and the Izotope Ozone Algorithms. I call bull. |
|
|
| rulzz |
| quote: | Originally posted by Morvan
On GearSlutz they boast themselves with being able to distinguish the Voxengo and the Izotope Ozone Algorithms. I call bull. |
there are people who indeed can do that |
|
|
| Magnus |
| Awesome and informative post Derivative thanks! |
|
|
| Derivative |
| quote: | Originally posted by Morvan
On GearSlutz they boast themselves with being able to distinguish the Voxengo and the Izotope Ozone Algorithms. I call bull. |
Hrmmm. Well Gearslutz can be a very useful place but like any internet forum theres alot of voodoo witchcraft and alot of people that misinterpret the science of digital audio. Occasionally someone like Paul Frindle or Bob Katz or Dan Lavry will set the record straight but for every one of those guys theres a hundred other people spreading misinformation. You can hear differences in dither noise shape algorithms if you turn up the volume of the dither so its really loud and then switch between algorithms. But turning up dither so you can hear it defeats the point of dither.
About 32 bit (float). I don't quite know how it works exactly so take the following with scepticism. I was under the impression that you have a 1 bit sign, 23 bits less significant bits and an 8 bit exponent. Or it could be 7 bit exponent, 24 bit mantissa and 1 bit sign. I forget exactly.
It has the same precision as 24 fixed bits. The exponent is for scaling amplitude in post processing if the signal processor can work internally at that precision. I think the exponent lets you destructively edit audio with 24 bits of precision and you can scale the amplitude back up to what it was before the edit because of the exponent. But don't quote me on that.
If its not going to crush your computer, eat up all your RAM and use up all your hard drive space with huge 32 bit IEEE renders then work at 32 bit float if you can. Otherwise I'm not sure it makes a whole lot of difference in terms of a good mix vs a bad mix. A good mix is going to sound great whether you work with 16 fixed bits, 24 fixed bits or 32 bit float. Whereas a bad mix is going to sound regardless.
If you want to get the proper facts I recommend going to Lavry Engineering website or going here: http://recforums.prosoundweb.com/index.php/f/38/0/
Dan Lavry is one of the good guys in this industry who takes the time to explain the concepts. Hes not out to make a buck with minsinformation and hes a true pioneer in the field of converter design. I haven't said anything that he hasn't said years before and better so if you have the time its worth a read. Bear in mind that its well into the realm of engineering and whilst that forum is good at simplifying the concepts, you probably won't find it very useful as a musician. Mainly its of interest to engineering types and people who design gear. It should all be foolproof in the sense that the end user shouldn't be expected to learn engineering terms to make music. |
|
|
| tehlord |
Great post again D :)
I'm going to file dithering under 'don't worry about it' for now. |
|
|
|
|
