fittingroom | browser letterfitting experiment | Recommender System library

It's a very, very experimental tool for type designers who want to play with machine-assisted letterfitting. Letterfitting is the process of adjusting the distance between individual letters. Given the number of letter pairings there are, this is a tedious process – but it's an important one, because badly fitted fonts look drunk. Conventional wisdom says space first, kern later, meaning: first set up left and right side bearings around each letter to make everything look okay, then make adjustments to individual pairs (the so-called kerns). The distinction between sidebearings and kerns has two good reasons: first, it's how metal type used to work, second, it's a lot easier than fitting every pair individually (plus, it makes for smaller files, duh). Because this distinction is so practical, it's easy to overlook that it's entirely artifical and arbitrary: what matters, really, are only the distances between pairs. What if, instead of worrying about sidebearings and kerns (and how changing one will mess up ten others elsewhere), you could simply drag letters around, and your computer could figure out what you mean, and apply changes throughout the font to reflect those stylistic choices? And at the end of the process, a magical algorithm hands you a list of sidebearings and kerning values? Wouldn't that be awesome?. Okay. We'll have to first identify letters that are similar to one another. For example, if pairings with o have certain letter distances, then those for pairings with q are probably similar, because they're both round. People have thought of this before: Frank Blokland and Pablo Impallari have, for example; and the people who thought up OpenType kerning classes have, too. But not all letters can be placed into neat groups. What's more, groups may differ from typeface to typeface. So let's embrace this haziness of real life, and resort to math. We'll use a technique called nonnegative matrix factorization. It generates for us a set of grayscale images which, kinda-sorta, represent proto-letters such that all actual letter shapes can be considered a weighted average of these images. In machine-learning speak, the images are called components or sometimes features (I'll call them the latter for now) and the weighting coefficients are, well, weights. If we limit ourselves to two or three of such features, they will look like blurry grey blobs; with more, you'll clearly recognize them as representation of letter groups (round, straight, etc.). When you use as many features as letters, each feature will be exactly one letter. You get the idea. Of course, because there are fewer features than letters, the weighted averages will never be completely faithful reproductions of the letters. But that's a fair price to pay.