Why adding noise to your digital recording makes it sound better
Here is a quick five-point guide to analog-to-digital conversion...
- Sample the voltage of the signal 44,100 times a second. 'Sampling' in this sense means grab the instantaneous voltage at each 44,100th of a second time interval.
- Hold that voltage for a short time, approaching the next time interval (these two processes are known together as sample and hold
- Measure the voltage very accurately.
- Only 65,536 voltage intervals are allowable so decide which one the measured voltage is closest to - this process is called quantization.
- Give the sample a number using the binary system - for example 1000101001010011
- That's it - analog to digital conversion! Sorry, that makes six points.
That was easy. Well, easy to describe briefly, more difficult to do well. That's why analog-to-digital converters often sound different to each other. To do all of this accurately within the short time spans allowed, and to very small increments in voltage, is a tough task.
There is another problem. Since the analog voltage input is infinitely variable, and only 65,536 steps are allowed (this figure relates to the 16-bit compact disc system, as does the 44.1 kHz sampling rate), then there is a certain 'rounding' error involved at each sample. The error can be as much as half a step, which might not seem much but it happens 44,100 times a second and it all adds up.
There are two outcomes of this. Firstly, the error has a random element, and randomness manifests itself as noise. This isn't that much difference from Brownian motion causing air molecules to vibrate randomly, causing acoustic noise that microphones can and do pick up; and the random motion of electrons causing noise in an electric signal.
However, much of the error in quantization is not random, it is related in some way to the signal. We don't call this noise anymore, we call it distortion.
It is a fact that digital distortion is gruesome. Far worse than the analog variety. It could well be described as being similar in awfulness to the distortion often heard on a poor FM radio transmission. That means bad.
Fortunately, there is a solution, and that is to add noise to the analog signal! Amazingly, adding noise of a level of one-third of a quantization interval magically cures the distortion. It also gives the system a consistent noise floor which is also desirable as the ear doesn't like noise that changes in level.
This noise signal is known as dither, or dither noise. Without it, digital audio as we know it wouldn't really work.
Dither has to be applied elsewhere in the audio signal chain, but in the analog-to-digital converter, it is done as a matter of course and the engineer doesn't need to know about it.
Oddly, dither also allows signals to be recorded that are lower in level than Level 1 out of the 65,536 allowable levels. Amazing but true - the dither signal will raise the level of very low signals so that they register and are recorded, rather than being ignored because they are below the lower limit of sensitivity.
Whoever would have thought that noise in audio would sometimes be useful?