Re: [ARSCLIST] Restoration Software

[Jon Noring <jon@xxxxxxxxxxx>]

Joe P. wrote:

> I would still go with Sonic Solutions No-Noise - especially the Type
> E interpolator. It is expensive, but the results are world class
> (Cedar is also very good)...

I can't say I'm an expert on the various top-end sound restoration
software and firmware -- I'm definitely not -- but I have experimented
with some in the past.

I was a close acquaintance of George Morrow, a Bay Area collector of
primarily 1920's and 30's dance and personality recordings (and a
well-known early pioneer in the personal computer industry) who sadly
passed away two years ago. George used Sonic Solutions for 78
restoration. His restoration work was quite good (which seemed to
slip, however, in the last couple years of his life as his health
deteriorated.) George believed that Sonic Solutions was superior to
Cedar, for what it is worth.

Anyway, in discussing 78 restoration with George, where pops/clicks/
crackle is the dominant form of noise to remove, that to do a good
job, handwork to search and destroy clicks is still a necessary step.
George found that if Sonic Solutions automated pop/click removal was
"stomped on" too hard to remove clicks, it left annoying audible
residual artifacts (I've also noticed this as well experimenting with
other similar tools on the market, such as Waves.) George found that
it was better to use these automated tools very gently to remove some
to most of the pops/clicks/crackle, then manually remove the remaining
pops/clicks one at a time.

This is where it gets interesting. George would listen to a one
second or so segment of the recording, discern any pops/clicks, then
look at the wave form of that time segment to locate, scanning by eye,
the location of the "glitches" he heard. He would then remove each
glitch, using Sonic Solutions to reconstruct the wave form by some
high-level, proprietary interpolation algorithm. After a while, George
got pretty good at finding glitches scanning the waveform by eye
without even having to listen; he'd show me where the glitches were in
a closeup of the waveform, most of which I could not even discern --
he'd remove each of them and the clicks were magically gone.

Clearly, the process used by George is fairly inefficient (I watched
him in action as I noted), particularly in locating the exact time of
each glitch. A while back I mentioned what I believe to be the "dream
tool" to allow for easy and precise hand removal of pops/clicks,
almost in machine-gun fashion. As far as I know (based on an
assessment I did a year ago), no one has yet built this tool, either
closed source or open source. It comprises two parts:

1) Easily locating the exact time and approximate width of pops/clicks
   using a spectrogram (frequency vs. time map.) In a spectrogram, each
   pop/click is very easily seen as a sharp vertical line which usually
   extends way past the highest frequency in the recording (if need be,
   it can be verified as a pop/click by listening). The exact time
   location and approximate width of each glitch is then transferred
   to a separate wave form view as part of step (2) -- this can be
   done, for example, by the user clicking on the exact spot in the
   spectrogram where the click appears.

2) When the glitch is clicked on in the spectrogram, it is cut out and
   and reconstructed in the waveform based on the time and approximate
   width as found from (1). Reconstruction is done using a high-level
   interpolation based on analyzing the spectral content on both sides
   of the glitch, maintaining phase as much as possible in the
   reconstruction. Note that this reconstruction will be done on the
   waveform (amplitude domain), not in the frequency domain used for
   step (1) (I won't explain here why I mention this.)

   I believe there is one published algorithm (and maybe more) to do
   the high-level reconstruction meeting the above requirements. This
   is an area requiring further research, but I believe it is not a
   show stopper (it is likely that Sonic Solutions, or whoever now
   owns their IP, will not reveal their algorithm.)


Since it's been a while since I last looked at this, it is entirely
possible someone has built (or is in the process of building) such a
tool. But if not, I'm hoping that someone will take the interest to
do so, preferably as an open source effort, maybe organized at
SourceForge. I will certainly join the effort, but I can't lead the
actual development effort because of my lack of programming
experience. (As an aside, I hope that someday all high-end audio
restoration software will become free, developed as part of open
source activities.) Of course, comments and criticism of this idea,
and the proposed open source effort, are welcome.

Jon Noring