"Mike Scott"
wrote in message
Jim Lesurf wrote:
In article , Rob
wrote:


That's really why I ask - I think. If there's more than
one way to downsample properly, I'm stuffed.

In principle 'downsampling' should be done 'properly'
and will then lead to a uniquely defined results - even
if done in various algorithmic ways.

That's not so.

If you are defining "uniquely defined" as being some precise bit pattern,
then I am forced to agree.

Downsampling always involves a reduction in Nyquist
frequency. It's necessary therefore to filter the input
to make sure frequencies above this are sufficiently
reduced. That filter can never be perfect, and there will
be various tradeoffs, involving extra loss of top-end,
in-band ripple and 'wrap-around' garbage from
insufficient rejection of higher-than-Nyquist signal.

That would be one of those things that is true - theoretically, but from an
audibility standpoint, is not true.

The big difference is how sophisticated we have become in terms of designing
and implementing digital filters.

It's all down to what the person doing it thought would
be best (by some arbitrary criterion), and there is no
unique or 'right' answer.

If computational resources are highly estensible, it is possible to product
digital filters with very nearly ideal phase and amplitude characteristics.

The realm of perceptual studies have also improved - we now know that the
ideal phase characteristic for the required brick wall filter is neither
linear phase nor minimum phase. However, we base that knowlege on
experiments done at Nyquist frequencies well below 20 KHz, because sonically
innocious downsampling to 22 Khz has been routinely availble at a reasonble
cost for nearly a decade.