I've just uploaded my scans of 412 Welte T98 ["Green"] rolls to the
IAMMP site! This can be found at its new URL,

http://www.pianorollmusic.org/rolldatabase.php?sortby=filedate&go= 

These are all Freiburg Welte rolls, a mixture of classical and dance
rolls from 1905 to the early 1930s. All of them embed the year of
issue and those meeting the U.S. public domain rules can be downloaded.

Over time the IAMMP site will progressively grant access to the
presently-blocked files. The vast majority of Welte is public domain
(PD), mostly now being over 100 years old. Their library of recordings
by famous pianists is superb, and many of the roll scans are of that
nature. Only a few of the later dance rolls are still protected.

All the MIDI files were made from reconstructed masters so the timing
is precise and skew is removed. They have simulated-expression created
using my own model of the Welte system. Paper acceleration allowed for
by adding a Tempo map; done that way the underlying roll matrix isn't
affected, it's simply played increasingly fast like a real player does.
The increase is calculated every 1/4-inch, an approximation but MIDI
doesn't allow for continuous increases. To be listened to via your
computer or a MIDI piano, I use the PianoTeq simulator.

Welte rolls can be very long, and have a great deal of lead and runout
paper, making them time-consuming to scan. Processing to get to some
form of the original 'master' is also time-consuming. The rolls are
actually 1-to-1 copies of playable rolls (which Welte called the
'master'), done at a very fine step resolution of about 60 rows per
inch. Bridging was imposed at perforation time, so every batch is
slightly different because the bridges are a little different as is the
reading of the master.

The very fine steps are a challenge to scan, but the horizontal lines
of bridges across the roll, and the constant lengths of the bridges,
provide a firm anchoring point for the software. In particular, they
give a base to de-skew the roll. It must be said that the test patterns
cut at the end of many rolls rather suggest that Welte's cutting wasn't
too precise in avoiding skew, and which might in part explain the
differing results obtained from individual copies of the same
performance. Everybody else used sprocketed master stencils that
avoided such problems.

There might be some improvement from statistically merging multiple
scans to find a best-estimate position of the note starts and get back
to the original performance as best possible, but given the scarcity of
these rolls that's not a practical option. It might be worth trying a
couple of time to see if informs anything about the systematic errors
in the production process.

Julian Dyer