0
$\begingroup$

To record sound and transform it to digital form, audio signals are recorded and its pressure levels are sampled and quantised.

Why does quantisation noise sound like white noise?

$\endgroup$
  • $\begingroup$ because your hearing is not an adequate measurement instrument to tell correlated noise from white noise easily. $\endgroup$ – Marcus Müller Apr 21 '17 at 18:39
  • 4
    $\begingroup$ It also depends on the number of quantization bits. For very low numbers of quantization bits, you can actually hear that it doesn't sound like white noise. $\endgroup$ – Maximilian Matthé Apr 21 '17 at 18:45
  • 1
    $\begingroup$ Given a waveform transitions multiple quantization levels per sample, and if the waveform is not correlated to the sampling clock, then the remainder is independent, uncorrelated from the remainder in the previous sample, which by definition is white noise. (although uniform white noise in this case) I think it would be interesting if someone could show how quantization noise relates to modulo number theory in the generation of random sequences based on what I described (and how a sine wave would compare to a ramp in the sequence generation that results). $\endgroup$ – Dan Boschen Apr 21 '17 at 18:51
  • 2
    $\begingroup$ @MaximilianMatthé Wow, you just happened to have a whole blog post just for this question!! $\endgroup$ – Dan Boschen Apr 21 '17 at 18:53
  • 1
    $\begingroup$ To people are answering in comments: short answer are OK. $\endgroup$ – Olli Niemitalo Apr 22 '17 at 8:21
2
$\begingroup$

It often doesn't sound like white noise. The quantization noise of a sine wave is fairly harmonic with a clearly detectable pitch. The quantization of a pulse train sounds like a pulse train.

If the quantization has enough bits, and if the signal is reasonably stationary in time and reasonably dense in the frequency domain, the quantization noise will be more or less white. For music quantized at 16 bits, that's often the case. However the quantization noise will always be correlated with the original signal which can produce undesirable artifacts. Hence commercial audio production uses dithering and noise shaping and decrease the audibility of quantization.

$\endgroup$
1
$\begingroup$

(following Olli's suggestion, posting my short answer): As long as the waveform transitions through multiple quantization levels per sample, and if the waveform is not correlated to the sampling clock, then the remainder to the closest sample boundary (which is the quantization error) is independent, uncorrelated from the remainder in the previous sample, which by definition is white noise. (Although uniform white noise in this case if there is an equally likely chance that the actual signal can be anywhere between two quantization boundaries). Any correlation between the signal and the sampling clock will result in pattern repetition of the error signal, and pattern repetition results in discrete frequency correlations that we call "spurs".

Please see the good work Olli Niemitalo has detailed on quantization noise here:

Quantization Noise for Coherent Sampling - Phase Noise?

$\endgroup$
1
$\begingroup$

It does not! The rounding errors introduced by quantisation are dependent on the audio, and when the input waveform has a repeating pattern, the quantisation errors will have a repeating pattern as well.

This is why we add dither: extra random noise before rounding. If our random noise follows the right distribution, we can ensure that the rounding error for each sample is (1) not correlated with errors at other samples, and (2) has constant mean and variance. When the rounding errors have these properties we perceive it as white noise.

Here is a page with some audio demos so you can hear non-white-noise quantisation errors, and how dither can improve the sound: http://www.earlevel.com/main/2010/11/07/the-sound-of-dither/

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.