[ARSCLIST] Frequency-space visualization (was: Noise reduction software for 78rpm)

[Jon Noring <jon@xxxxxxxxxxx>]

Divine Art Record Company wrote:

> The biggest difficulty I find with Dcart is locating small clicks on the
> waveform display, as the zoom function seems to shift the display sideways.

I'm still amazed that most manual click removal tools don't employ
wavelet-derived frequency-time-space visualization to locate the small
clicks left over after the automatic declicking/decrackling tools do
their magic (certainly, the settings of such tools can be configured
to effectively remove 100% of the impulses, but then, as Divine Art
noted, the sound is horribly distorted as a result.) Small audible
clicks which are hard to locate and visualize in the waveform stick
out like sore thumbs in frequency-space (wavelet mode, not FFT mode)
and can be trivially selected with the "click" (pun not intended) of
a mouse, almost machine-gun like in fashion -- like shooting fish in
a barrel, so to say.

Pristine Sounds 2000 is a wavelet-based frequency-space editor I've
experimented with for removing clicks in 78 material (I can manually
remove 10 to 20 in a *minute*, almost machine-gun like fashion.)
Unfortunately, it has the fatal flaw that even though it *exactly*
determines the location of the small clicks, the tool it uses to
reconstruct them does it in frequency-space (rather than just marking
the exact time and directly editing the waveform by "cut and
reconstruct" of the samples) -- thus each edit changes the waveform
far outside the click area (well over a whopping 200 samples on both
sides of the click!) If there are just one or two clicks in a tenth of
a second timeframe, the repair is not too noticeable, but if there are
several in such a time frame (and this is common in the more noisy
78s), then their combined removal results in a very noticeable
unevenness ("pumping") in the background hiss, which is very annoying
and almost as bad as the clicks themselves (the reconstructed music
sounds great, though -- but variable/pumping background hiss *is*
quite annoying.)

"Cut and reconstruct" should be to simply cut out the several samples
in the waveform itself where the click is located (which can be from
several to a few dozen samples wide) and then reconstruct the "gap"
based on high-order frequency analysis of BOTH sides of the gap thus
assuring the reconstructed portion has the correct spectral content
and time synchronization/harmonization -- no samples outside of the
removed gap should be changed. There are algorithms to do this
high-order reconstruction (i.e., Sonic Solutions employs such an
algorithm, and manual click removal and reconstruction using SS
produces superb results as George Morrow's exemplary 78 restoration
work shows -- unfortunately SS does not have frequency-space
visualization to quickly locate the clicks thus one has to laboriously
scan the waveform with squinty eyes to try to find the click.) As an
aside, what should *never* be done to "fill in the gap" is simple
interpolation, either a straight line connecting the ends of the gap
or simply zeroing out the samples -- this in effect replaces one type
of noise with another type of noise, and I can audibly hear such
"linear" fixes -- they are almost as annoying as the original click.

Anyway, I've been in touch with the developer of Pristine Sounds, but
at the time I talked with him, he was totally unmoved by my suggestion
since the intention of his tool is to do complex frequency-space
editing for advanced waveform processing (such as filtering out
certain musical instruments, etc.) and he simply saw no "market" for
optimizing it as I describe above for 78 source click removal.
Nevertheless, PS 2000 demonstrates how such a killer manual 78 click
removal tool would work. It would revolutionize how 78 material is
restored, and allow one to fully declick the more noisy 78's in a
fraction of the time it would take to do using the traditional method
of trying to locate the clicks by scanning the waveform (as an aside,
it means that one would be able to more gently use the automatic
declickers and decracklers, thus resulting in even less distortion of
the wanted sound.) I oftentimes find with full manual declicking that
final denoising is not necessary, and even when it is called for, one
can use it much more gently. I believe the key to getting great results
out of older 78 restoration is complete (and properly done) *manual*
declicking. Removing 300 to 500 clicks in a recording is not an issue
when one can machine-gun them out in a matter of an hour or so --
wavelet-based frequency-space visualization allows one to do it.

If the Diamond Cut people are reading this, hopefully it will spur
them on to investigate the use of wavelet-based frequency-space
visualization for manual click removal and reconstruction. I'd do it
myself if I could, but my programming skills are zilch and I don't
have enough resources, or time, at the moment to setup a venture and
hire the programmers to build the tool (I'm working on other ventures
at the moment.)

Note, CoolEditPro and such tools do have frequency-space visualization
modes, but to get them to work properly for manual click removal is
difficult, and even when setup properly they do not even *remotely* come
close to the convenience, speed and ease demonstrated by Pristine Sounds
2000 (and anyway, both their frequency-space visualization and wave
reconstruction algoritms suck royally.) (As just noted, too bad PS
2000 does not do the right thing when it comes to the removal and
reconstruction of clicks.)

Jon