Re: [Audacity-devel] Find Zero Crossing bug

[Al D. <bus...@gm...>]

> Hi Al
>
> FWIW, most users I've ever seen commenting on zero crossings expect it to
> simply find the nearest zero crossing. Maybe what we're trying to do is
> better, and clearly we're making better decisions than before in many
> cases. Still, I picked a small part of a stereo music track at random
> (below) and I'm not sure the result is intuitive to (my) naked eye.
>

I'm not sure that what we're doing is better than simply finding the first 
ascending zero-crossing...

> Here's a region before finding zero crossings, the "nearest" zero crossing
> to the end is obviously at about 40.8486:
> http://www.gaclrecords.org.uk/zero_crossing_before.png
>
> But we find the end crossing at 40.84755:
> http://www.gaclrecords.org.uk/zero_crossing_after.png
>
> The 40.8486 crossing has the left channel sample dead on zero, and the
> right channel has samples touching zero, above and below.  The 40.84755
> crossing has the samples in the left channel above and below zero, and the
> right channel has one touching zero (but not dead on it). So just on those
> immediate samples, it seems we ought to go to 40.8486.
>
> Now I know you are doing something a bit different (taking weighted
> averages of nearby samples) but I'm not sure this looks convincing. But in
> your defence, if I select the left channel only, the end crossing goes to
> 40.8486, where the sample clearly is closer to zero than the one at
> 40.84755. So am I being over-concerned abut the samples closest in terms of
> distance to the crossing, or are we spreading the average across too wide a
> field?
>
> On another topic, I recall several complainants have suggested we should
> always reduce the size of the region when finding crossings, the "loss"
> being better than including parts that were not asked to be part of the
> selection. Just throwing that in the ring. Just to look at Richard's
> example, if we select 0.2325 - 0.2625 the region expands (0.232 to 0.2635).
>
>

Or maybe we should simply find the nearest ascending zero-crossing that makes 
the selection size smaller.

I have been working under the assumption that the existing code is as it is 
for a good reason.  But it sure does produce some odd results, even with the 
patches I've added.  At any rate, samples adjacent to a zero-crossing but far 
enough from zero to pop are likely to pop in their original context anyway, 
right?  I'll try the approach of just finding the nearest zero-crossings that 
shrink the selection tomorrow.

 - Al