A couple years ago I experimented with this idea using material available
from the Great 78 Project, starting with groups of four transfers made
simultaneously using different styli on one of George Blood's four-tonearm
turntables. Even in that idealized case, synchronization was the Achilles
heel of the scheme. I wrote some code that compared the four transfers
using cross-correlation, worked out point-by-point timing differences down
to some split-second level (I don't recall the exact details), and then
resampled all four sources to an averaged time base, with some smoothing
applied on the assumption that each of the four deviation curves would be a
sinusoid. As I recall, I was able to get rid of the *periodic*
comb-filtering effect pretty well, but there still seemed to be some
similar distortion unfolding more gradually, if I compared (say) beginnings
and ends, with conspicuous long-term shifts in the quality of the surface
noise. I didn't keep any of the audio results (since I figured I could
just re-run the code), but I did create graphs of the timing differences
and included one in a blog post I wrote about the Berthold Laufer stereo
reconstructions. Here's a link to just the graph (the vertical scale is in
samples at 96 kHz):
And that's what I got from four transfers made from the same pressing on
the same turntable at the same time!
I tried applying the same code to transfers of two or three pressings of
the same matrix (whichever take of Arthur Collins' "The Preacher and the
Bear" was used for double-faced Victor pressings) just to see what would
happen. As I recall, the results were pretty ghastly -- that cyclical
comb-filtering came back with a vengeance. I think one issue was that
differences in impulse noise from varying groove damage threw off the
cross-correlation. I've always hoped to revisit this to see if some kind
of impulse-noise mitigation before cross-correlation would help, plus
comparing even smaller segments.
This is an intriguing idea, and it seems to hold a lot of promise, but it
would take some serious R&D to get it just right. Unfortunately, getting
it *almost* right is no good if the goal is for the result to compete with
an ordinary single transfer. Combining this approach with optical imaging
could be interesting -- it might be easier to align images of multiple
pressings than to synchronize audio transfers of them.
On Sun, Dec 20, 2020 at 12:21 AM Tim Gillett <[log in to unmask]>
> Hi all,
> I've familiar with the principle of summing two audio recordings of
> the same programme where the wanted programmes are time coherent but
> the unwanted background noises arent, making it possible in theory to
> realize a 3db reduction of the background noise. I've used it on
> cassette and 1/4 track "stereo" recordings which are actually dual
> mono, although time alignment can be a little tricky and an "azimuth"
> tool can really help.
> Earlier 78 RPM recordings are understandably noisy especially as I
> believe was customary, the original metal parts were usually recycled
> for cost reasons so all transfer engineers have to work with may be
> the best shellac consumer pressings that have survived.
> The other day while browsing I chanced upon a seemingly mint
> condition 10" 78 shellac record of Eileen Joyce performing
> "Berceuse" in 1939. Here's a modern CD release of it.
> Quite noisy and limited bandwidth, even for 1939 I thought. I
> cleaned and transferred the disc I'd found and after declicking
> compared the sound to the version linked to. The two versions sounded
> close. Signal to noise was about the same.
> I believe that compared to vinyl, shellac was a relatively noise
> medium so it occured to me that assuming the stamper had less noise
> than the shellac copies, it should be possible to time align the audio
> from the two discs, sum to mono and yield a potential background
> noise reduction of 3db. Not huge but with an already noisy recording,
> not to be sneezed at!
> So far I've not had much success. I'm having problems trying to
> accurately time align my disc transfer with the commercial release.
> I can get them within maybe a few milliseconds but it's not enough.
> The two versions drift in and out of sync causing comb filtering.
> My next step might be to find another mint disc of the same
> performance and make my own transfer of both on the same gear under
> the same conditions so at least there is a fighting chance of an
> accurate alignment. Maybe a better turntable with more precise speed
> Has anyone tried this technique with 78 RPM shellac records or have
> any comments?
> Cheers Tim.
> Email sent using Optus Webmail