I have spent some time working on transcoding between the different
reproducing piano systems and this inevitably leads to investigating
computer generation of Ampico expression.

The technique which seems to work best is based on a two-pass approach.
The first pass averages the playing power "lows" over a given interval
-- say, 2 seconds, but subject to much experiment.  This is used to
generate the crescendo data.

The second pass then adds Intensities to try and match the original,
note by note.  It is possible to take this approach much further by
iterating the steps and allowing the computer to adjust the crescendo
base and then try again.  After a few runs a surprisingly close match
can be obtained.

The real area of judgment lies in deciding how "busy" one is allowed
to keep the Ampico mechanism.  Note that there is no "music" in this
technique.  In my view if one can make each note sound at the velocity
it was recorded at then all is accomplished.  I don't go along with
those who argue that the analogue element in Ampico expression is
musically important.

Of course, the whole process only scratches at the surface.  It might
be possible to match Ampico expression to a MIDI file after a fashion.
How accurate was the MIDI file ?  How consistent are the Ampico pianos
which will be interpreting this expression ?

These are the real issues which will be-devil computer analysis of
piano roll music for ever.  Having watched the ludicrous quantities
of brown and sticky which has been flying in Wayne Stahnke's direction,
and knowing that I am light-years behind him, I have backed away from
the public arena for good.  I get hours of pleasure from my computer
transcriptions of reproducing rolls and that's what matters to me.

Richard Stibbons
Cromer, England