[time-nuts] Oh dear

Dan Mills dmills at exponent.myzen.co.uk
Mon May 7 18:37:46 UTC 2012


On Mon, 2012-05-07 at 18:15 +0000, Poul-Henning Kamp wrote:

> 
> We must start out by defining the acceptable level of total distortion,
> if we choose 0.5% then we need 200 digital levels, roughly 8 of
> your 16 bits for the signal.
> 
> That gives you a headroom of 7 bits (leaving one for the sign) and
> that gives you 42 dB of S/N.

Not true in a correctly dithered quantizer  (And they almost all are
these days)... 

This is counter intuitive, but adding 1 LSB of uncorrelated noise having
the correct statistical properties (Triangular probability distribution)
has the effect of completely linearising the conversion process at the
cost of adding about 3dB of noise to the system.
With the noise added you can hear narrow tones well below the wideband
noise floor.  
 
In a correctly dithered system the broadband noise floor is the only
thing determined by the word length, and narrow band signals can be
resolved to well below the noise floor.  

Further, as the statistical properties of the noise are not all that
tightly coupled to its frequency domain properties, it is possible to
filter the noise to move most of the energy away from the regions where
the ear is most sensitive. 

16 bits is actually fine as a distribution format, where is shows up as
a little short is as a capture format as at capture time you need
headroom to ensure nothing unexpected causes clipping, but once you are
done with the processing it is trivial to strip the headroom out and
dither down to 16 bits.

This discussion would be better over on the Pro audio list rather then
time nuts. 

73, Dan.





More information about the Time-nuts_lists.febo.com mailing list