Re: [Audacity-devel] Reassigned spectrograms

[James C. <cr...@in...>]

Spectral reassignment is in for 2.1.2.

It is a super feature, and 'local' in that bugs in it are unlikely to 
affect anything else.  Its bugs are mostly likely to be around quality 
of its own display, i.e. not reassigning as finely as we might like.  It 
was being worked on already in 2.1.1 rather than being completely new.


The track menu changes now give us something better than we had in 
2.1.1, and (as RM) that's good enough to justify making them.  As me, I 
think there is still something not as good as it should be in how we 
select track options, but I can't put my finger on it, and it does not 
scream enough to me to worry about it for 2.1.2.


I think Steve you are right to worry.  We have a tremendous amount of 
technical debt caused by the wx change.  3P1s, and 8 DEVEL-FIX-MADE P1s 
which are pending confirmation that they have been fixed (without 
breaking something else important).  Any changes as big as spectral 
reassignment intended for 2.1.2 from here on need to ask "Is it OK?" on 
this list and the answer is probably no.  Minor features/improvements 
are definitely OK and in particular picks for these ones from Gale's 
wishlist:

  * Option to turn off backwards scrubbing OR a modified click that
    automatically plays forward from the seek point. This is to reduce
    un-necessary RSI after seek.
  * Some TLC for those who do large amount of daily export and import:
      o Bug 550 – Add Directories Preference to export to directory the
        file came from
        <http://bugzilla.audacityteam.org/show_bug.cgi?id=550>
      o Related to above, easier one-step overwrite of the same file you
        imported.
      o Bug 549 – Add missing export formats to Chains
        <http://bugzilla.audacityteam.org/show_bug.cgi?id=549>andChains
        export format parameters and sample rate
        <http://wiki.audacityteam.org/wiki/Feature_Requests#Improvements_to_the_.22Chains.22_.28batch_processing.29_feature>
      o Bug 551 – Export Multiple: Pass through common metadata to next
        window <http://bugzilla.audacityteam.org/show_bug.cgi?id=551>
      o Support per-track metadata so that if you import multiple files
        you can export multiple using each file's metadata instead of
        export all files with only the last file's metadata
      o Bug 701 - Embedded album art and lyrics discarded
        <http://bugzilla.audacityteam.org/show_bug.cgi?id=701>.


I would like to see some more work on noise reduction to make it 
better/easier to use.  I would like Paul to set a colour for spectral 
selection that he is happy with to replace the yellow. We are leaving 
'rough edges'.  We do have until 30th August before Semifreddo.

Steve, does this answer your concerns?

--James.


On 18/08/2015 11:39, Steve the Fiddle wrote:
> I'm just wondering why we are seeing so many new features introduced
> for the 2.1.2 release. I thought that this release was about updating
> to Wx3. My concern with so many changes in this release is that either
> quality suffers or the release is seriously delayed. Perhaps James can
> comment and reassure us that it is all in hand?
>
> Steve
>
> On 17 August 2015 at 15:40, Paul Licameli <pau...@gm...> wrote:
>> I don't think it's "experimental" any more, I think I got all the math right
>> now.
>>
>> There are opportunities for more efficiency.  It is true that caching of
>> results leaves something to be desired for some changes of view, but that is
>> already true of spectrograms generally.  There is no caching of any
>> spectrograms for changes of zoom level.  There is some caching only for
>> scrolling left and right without changing the zoom, and that much won't work
>> now with reassignment.
>>
>> I consider any such effort for performance a future project.
>>
>> PRL
>>
>>
>> On Mon, Aug 17, 2015 at 10:28 AM, Steve the Fiddle
>> <ste...@gm...> wrote:
>>> Is this an experimental feature?
>>>
>>> Steve
>>>
>>> On 17 August 2015 at 15:18, Paul Licameli <pau...@gm...> wrote:
>>>> As of
>>>>
>>>> https://github.com/audacity/audacity/commit/3ddbbd375d6963838b91fe2f54c91df96d90e79a
>>>>
>>>> Now reassignment is a third algorithm choice in spectrograms preferences
>>>> and
>>>> view settings.
>>>>
>>>> This is an algorithm that sharpens the results of a spectrogram.  It
>>>> works
>>>> by computing two extra FFTs using different windowing functions.  The
>>>> contributions that the usual FFT makes to pixels on the screen, are in
>>>> effect moved slightly using results of the other FFTs.  This can
>>>> counteract
>>>> both the "time smearing" and "frequency smearing" effects.
>>>>
>>>> But not only is this spectrogram about three times as costly to compute
>>>> -- I
>>>> also had to disable the caching that allows horizontal scrolling of
>>>> spectrograms to go faster.  This sort of caching is important for
>>>> scroll-scrub with spectrogram on.
>>>>
>>>> The zero padding of spectrograms that I implemented in the last version
>>>> was
>>>> meant to accompany this.
>>>>
>>>> Try it out with tones and chirps -- you will find just a thin bright
>>>> line,
>>>> even if your window size is small!
>>>>
>>>> Try it with an impulse -- you will find a bright vertical line, even if
>>>> your
>>>> window size is large!
>>>>
>>>> Try it with real world examples -- and it is true there are strange
>>>> artifacts from "cross terms" if you have two or more sources of sound,
>>>> or
>>>> inharmonic sounds, which make a sort of gauzy pattern as in this
>>>> article.
>>>> https://en.wikipedia.org/wiki/Reassignment_method
>>>>
>>>>
>>>> PRL