[music-dsp] Declipping

[robert bristow-johnson]

i deleted my name from the subject line.  i see myself as a participant, not
the subject.

on 02/21/2005 13:16, Knut-Inge.Hvidsten at ffi.no at Knut-Inge.Hvidsten at ffi.no
wrote:

> Information is lost. In theory the clipped samples COULD have been any value
> between (max value + 1) and inf. At best what can be gained is masking the
> error by assuming something about the source (or destination).

yup.

> clipping points could be identified manually or automatically if some treshold
> is reached (typically full-scale sample), or some combination of both
> (manually enter a "zone" that seems to have one or more clips).

yup.

> Assuming that music is usually slowly varying (lp), what about extending the
> value, 1st and 2nd derivate from each end? if they cross at a sharp angle,
> round the top? :-)I have no clue what it will sound like, but I guess that the
> information content lost in a clip usually is low?

extending the low-order derivatives makes sense when the assumption is
"usually slowly varying".  that assumption is similar (if not equivalent) to
lacking high-frequency energy.  the clipping causes discontinuity for the
first (and higher) derivatives which would introduce high frequency content.
that's what formed the basis of my *guess* for an error metric.

-- 

r b-j                  rbj at audioimagination.com

"Imagination is more important than knowledge."