On 2006-05-19, Serge Auckland wrote:
Ian Iveson wrote:

The DSP solution would appear to be preferable. However, presumably there are
rounding errors. That is, if I divide every word by a constant, each result must
be rounded to the nearest step, and this rounding error is not linear wrt the
audio signal. I assume there is a name for this kind of error? How significant
is it?

The DSP solutions I'm familiar with operate either floating point or 32
or 48 bit internal which means that when the output is finally reduced
to 16 or 24 bit the errors from the DSP calculations are minimised. I
don't know of a specific name for the errors resulting from DSP operation.

I guess a digital volume control can just be a single fixed-point
multiply.

If you have (for example) a 16-bit coefficient and multiply by the 16-bit
audio value you get a 32-bit result.

Growth in significant bits is typical of DSP actions such as add and
multiply. If you just truncate the result back to 16 bits you do indeed
find that the quantization error you introduce is correlated with the
signal. It is significant and sounds bad at low levels.

In a good DSP (volume control) the remedy is to keep all significant
bits throughout the entire operation (or at least keep enough) and add
dither noise before you finally truncate back to the desired word size.
The random dither de-correlates the quantization error from the signal.

I assume that's what good DSP volume controls do.

--
John Phillips