Charles M. Chen's Autokern is a command-line Python tool released as part of Chen's TypeFacet software collection. It reads in UFO font files and takes a set of user-supplied arguments, then outputs a modified UFO file with kerning values and, optionally, adjusted side-bearings.
Autokern is run from the command line and the advised usage is to execute multiple runs. First, one should do a set of test runs on small glyph sets to determine the values of font-wide parameters. Then one can run Autokern on the entire font.
###Determining parameters
First, one runs Autokern with (essentially) all zeros; this sets the
min-distance-ems, max-distance-ems, max-x-extrema-overlap-ems
and intrusion-tolerance-ems parameters of the font to 0. After this
step, one can inspect the output UFO file to note which glyph pairs
may be problematic.
Second, one runs Autokern again (on the same input UFO) to set the
min-distance-ems parameter to some non-zero value. Selecting the
appropriate value is up to the user based on trial runs and inspecting
the results. The parameter is meant to reflect the minimum spacing
between two glyphs that come as close as possible to touching. For
example, the pairs "TT", "vv", or "rf". During this run,
max-distance-ems is also set to the same value, although it will
change later.
Third, one runs Autokern again, this time to set the
max-distance-ems parameter. Again the correct value is up to the
user to determine by inspection; it is meant to be the maximum space
allowed between two glyphs. For Latin, the example pairs recommended
are "NN" and "||". For serif faces, max-distance-ems is meant to
reflect the distance between stems.
Fourth, one runs Autokern to set the intrusion-tolerance-ems
parameter, which defines the maximum amount that glyphs can "intrude"
into the space between them. The examples indicate that
such intrusion amounts to the distance that an extrema point can be
kerned into unkerned space of the sidebearings between a kerning
pair. Furthermore, regardless of what the intrusion-tolerance-ems
setting is, no pair of glyphs will be kerned closer than
min-distance-ems, preventing collisions. The example pairs
recommended are "AA" and "OO".
Fifth, one runs Autokern to set the max-x-extrema-overlap-ems
parameter. This parameter sets the amount of (non-colliding) overlap
allowed between the extrema of a kerning pair. For example, the
inner extrema of the classic "AV" and "VA" pairs are likely to overlap
each other horizontally, even though the contours do not touch.
###Kerning an entire font
After the parameter-setting tests described above, Autokern can be run
on an entire font. The precision-ems parameter sets the smallest
step size that Autokern will use when making adjustments. Selecting
too small of a value will result in long run times.
There are several other arguments that one can make use of at this stage for better control of the output. In particular, arguments can be used to limit kerning adjustments to a specific class of glyphs, so that different rules can be applied to different scripts, cases, and glyph types. Autokern can also produce helpful log files, including HTML output that may be easier to work with than flat numeric data.
The Autokern algorithm depends heavily on the parameters detemined in the preceding steps.
For each kerning pair, two "facade profiles" are constructed for each
glyph: one pair by "inflating" each glyph a distance of
min-distance-ems and one pair by inflating each glyph a distance of
max-distance-ems.
The min-distance profiles are moved together until they touch, as are
the max-distance profiles. This determines the "minimum advance"
and "maximum advance" values for the kerning pair.
Next, the max-distance profiles are adjusted to determine the
"intruding advance", which is the distance adjustment that results in the
most equally balanced "intrusion" amount for the two glyphs—that
is, each glyph intrudes by a close-to-equal amount, without exceeding the
intrusion tolerance parameter.
The intruding advance is found via binary search on the interval [0, int(ceiling(2.0 * max(pair_intrusion_tolerance, pair_max_distance)))]. At each step, the algorithm takes the horizontal distance between the facade profiles at several y-axis values, looking for collisions between the profiles. If a collision is detected, the next iteration moves to the "far" side of the remaining search interval.
The final kern is determined by summing the intruding advance (or the
"minimum advance" value, if that is greater than the intruding
advance), any tracking-ems value provided as a parameter, the
x-extrema-overlap-scaling parameter, and the greater of the
x-extrema-overlap and max-x-extrema-overlap parameters.
The facade profiles used in the collision-detection and intrusion
steps are of reduced precision (in comparion to the font's internal
point system); the precision used is derived from the precision-ems
parameter.
The algorithm can also be tweaked with the
ignore-x-extrema-overlap-outside-ascender parameter, which
limits the spacing-adjustment process to measurements within the x-height.
Bluntly speaking, Autokern makes "common sense" kerning calculations based on what it assumes to be a properly letterspaced input font. Which is to say that it attempts to make unobtrusive changes.
As far as the critical "intruding advance" step is concerned, the algorithm measures the horizontal distance between the two glyphs at a number of y-axis sample heights, but the samples also take three specific conditions into account:
-
Large contiguous gaps at the top of bottom are discarded; this removes the effect (for example) of the left ascender in the pair "hh".
-
Large gaps in the middle of the height samples are de-emphasized by replacing them with the maximum protrusion value. This removes undue impact from the open bowl in "c".
-
Sample rows that exist solely because of the glyph-inflation process are discarded. This essentially chops the profiles off at the original glyph's vertical extrema.
Together with the pair-intrusion-tolerance parameter, these conditions
restrict the intrusion-advance search to commonly accepted rules about
Latin typefaces. That said, more specific rules might produce more
pleasing results. As implemented, the rules are somewhat ad-hoc. The
amount of "large contiguous gap" that triggers the "hh condition is
hard-coded to pair_max_distance * 0.5. Depending on the typeface,
that might trigger false positives for (say) "L" or"T".
Several of the other conditions in the intruding-advance step are similarly hard-coded. It is an open question whether these metrics produce the best results.
At a higher level, the glyph-inflation process used to create the facade profiles might bear closer examination as well. It "inflates" glyphs by a fixed amount, both horizontally and vertically, in an attempt to define a "bubble" into which adjacent glyphs should not intrude. But the fixed-radius inflation may not correspond well to how glyphs (particularly those with dense or complex contours) behave.
Finally, the usage of the Autokern tool itself requires the user to employ multiple, manual steps. Much of the functionality could be broken out into a reusable library to automate some of the parameter-determining stages or to work interactively.
The manual determination of the parameters (particularly
max-x-extrema-overlap-ems and intrusion-tolerance-ems) is a
crucial input to the final results. In a sense, since these
parameters are not automatically determined for a given input
typeface, the final output can only be considered partially automated.