Re: [Audacity-quality] bugs in 'Click Removal'

[Steve t. F. <ste...@gm...>]

Cc: to devel in case there is a developer interested in working on Click Repair.

but first, a short aside re. a comment to -quality in this thread:

On 13 December 2013 01:33, Vaughan Johnson <va...@au...> wrote:
<snip>

> And btw, there's a general principle that commands should be verbs, so
> e.g., that one and Silence Finder, should be 'Remove Clicks' and 'Find
> Silences' (or even 'Label Silences' or 'Mark Silences').

<snip>

I don't think that "general principle" applies to effects.
Some very common effects:
Reverb (noun)
Chorus (noun)
Phase (noun)
Distortion (noun)


Back to the issue of Click Repair -

I've been experimenting with various methods for repairing clicks, and
my results show that our "Click Repair" effect is at least doing one
thing right.

Having tested a number of methods, for short clicks, simple linear
interpolation across the click region is, according to my results, as
"good" or "better" than other methods. In this context, "good /
better" takes into account the effectiveness of reducing the
audibility of the click, and the amount of "damage" (loss of valid
audio information). Also, linear interpolation was the quickest
method. The repair with linear interpolation is rarely "perfect" but
with almost all real world cases the repair is virtually inaudible.

* An exception to the above is if a click is caused by slight
clipping. In this case, linear interpolation is ineffective, but imo
this case should really be handled by a "clipping repair" effect
rather than by a "Click Repair" effect.

For the tests I used a selection of clicks of up to 1 ms duration. For
clicks over 1 ms, LSAR interpolation was better (as used in the
"Repair" effect), but becomes unreasonably slow for repairs much over
2 ms. In my opinion, damage much longer than this is not really a
"click" (more like a "pop" or "noise") and cannot reasonably be
expected to be repaired by this type of effect. Methods tested
included various types of interpolation and filtering.

The primary weakness in the current click repair effect is the (overly
simple) detection algorithm.

I am currently working on a Nyquist implementation of a new algorithm
as a proof of concept, which looks like it will be much more
effective. However, it is likely to be slower than the current effect,
even if implemented in C++. (the Nyquist implementation will probably
be *extremely* slow). For this effect, how important is speed?

Steve