I decided to try Kern On with my latest typeface. After deleting all the small pairs (less than 5) and cleaning up some garbage (pairs for CALT glyphs) I still have over 7,500 pairs in each master. And the kern values are terrible for many pairs. Going through them one by one in Glyphs would probably take longer than just kerning the typeface in MetricsMachine. How are people dealing with this? Are people just crossing their fingers and hoping that Kern On did a good job?
Define models for some of those pairs to adjust the output?
The big problem I have found with previous auto-kerning is that even when it did a good job in terms of internal consistency of output, it didn’t kern anything like the way I would have kerned manually. So I either had to accept that or throw it out and kern manually. The iterative model aspect of KO is what I like, and why I think it is the best approach. Also, Tim is very responsive to issue reports.
Usually I ended up with having at least 100 models and some exceptions. After that I test the kerning on some sheets I made on indesign and web browser, refining every time going back and forth in the master file and in the test specimen. I have to be sure to exclude from the kerning engine a lot of glyphs and I tend to keep the kern table as small as possible (15kb to 25 kb). I also run a lots of scripts that helps me to check the result (kerncrasher, overkerned pairs, large pairs all of them from mekkablue library).
It is a long process and full of small mistakes. But at the end I will have a more than acceptable kern mostly extensive, saving a lot of time and my mental health.
At the end I accept some imperfections in my kern (at least in the less important pairs) it is the nature of automatic generated stuff, I always think that my eyes are much more reliable than a computer, but I'm not sure anymore if this is true or just my bias (I always thought computers are just fast and stupid).