The main improvement in Audiveris processing stems from
the use of filaments organized in a kind of grid.
This grid serves as a referential for the following tasks,
and provides an elegant approach for skewed or distorted
Several installation files have been published, all using
the 4.1beta core name.
This reflects the status of continuous development rather
than stable release of the software.
Filaments: They are long glyphs representing
the core of either horizontal or vertical lines
(staff lines candidates and barlines candidates
These filaments are formalized in natural splines,
which are sequences of Bézier curves with continuity
up to the second derivative.
Grid: The staff lines and barlines are
connected into a grid of sometimes rather wavy lines.
The grid itself is taken as the referential
for all the other glyphs, whatever the potential
skew or other distortion of the image, and thus
saving the need for any pre-processing.
Moreover, one can on demand easily build and save a
"dewarped" version of the initial image.
Scale: Additional key informations are
derived from run length histograms (jitter on staff
line thickness and spacing, typical beam height,
whether the image is music or not, ...).
Systems: The boundary between two consecutive
systems is now a broken line, resulting from the
incremental inclusion of glyphs into their nearest
Training: Besides full sheets taken as
training samples, the user can select a mode that
takes every manual assignment as a new training
family of compound symbols no longer exists, thanks
to an aggressive strategy in glyph split pattern.
OCR: We are still stuck to the old
Tesseract version (2.04).
The new Tesseract generation (3.x) has
been out for more than one year now but still lacks
a Java connection under Windows.
Time: All time values, such as offsets within
a measure, are computed using rational values,
which makes them independent of the score divisions