Script to subset fonts for the web

Smaller font files with only the glyphs I need

This post is archived. Opinions expressed herein may no longer represent my current views. Links, images and other media might not work as intended. Information may be out of date. For further questions contact me.

It is a good practice to optimise font files before self hosting them. Each font family may contain hundreds of glyphs that are not used anywhere on the website. This typically concerns the characters from scripts other than the one[s] you write in. Subsetting a font file removes those superfluous code points. It is the right way to reduce overall file size. For example, I only use glyphs from the Latin and Greek alphabets, as well as punctuation marks and numerals that are common to both.

Self hosted Google Fonts was a decent compromise

In the past, I would get fonts optimised for my use case from the Google Fonts project. It already offers font presets that include the supported Unicode subsets. You can download a ttf file or set thereof, then run command line tools to convert it to the woff and woff2 formats for use on the web. Or you can use a web app like the google webfonts helper which automates those steps for you.

The major downside with that method is the lack of control. You are limited to fonts that are hosted on Google’s platform. Some of the best free/libre typefaces are not available there, such as the DejaVu fonts, Mononoki, Hack, FiraGO, Iosevka, and so on.

Besides, the plethora of choice on that platform is largely an illusion. The available options are limited to a small group of fonts once you factor in the need for supporting multiple languages, bold, italics, etc. (most offerings are incomplete or at least not suited to my requirements).

Personally, I prefer the aforementioned free fonts. The problem is that they do not provide subsets or convenient built-in tools to reduce the Unicode coverage only to the set of needed glyphs. For my application, an optimised woff is about 50KB while woff2 falls to 30KB. Compare that to something around 1MB for Latin + Extended Latin + Greek and Coptic + Cyrillic…

This is where Google’s service has a clear advantage. It is why I would ultimately compromise on my demands, using self hosted variants of Google Fonts because I did not know how to subset fonts myself. And no, sending ~5MB of font data per initial page load was never something I wanted to do.

The “fonttools” package to the rescue

Subsetting fonts is no longer a hindrance. Today I discovered a package in the Debian repos which offers the means to subset a font from the command line. It is the fonttools collection of python programs.

sudo apt install fonttools

After figuring out the command I needed to execute, I wrote a small BASH script that automates the process for each ttf file in the present working directory. Below is the essence of the script in its current “alpha” version. I bundle it with my dotfiles under the “bin” directory, as this is something I might develop into a multi-purpose utility.


# This function accepts two arguments.  The first is the name of the ttf
# file without the file type extension.  The second is the desired
# output format (woff|woff2).  These are provided by the subsequent
# loop.
subset_font() {
    pyftsubset "$1.ttf" \
    U+0370-03CE,U+2010,U+2012-2014,U+2018-201F,U+2022-2027' \
    --layout-features='*' \
    --flavor="$2" \

# Loop through all ttf files in the present working directory and run
# the `subset_font` function defined above.
for i in $(find ./*.ttf | sed 's,\(\./\)\([a-zA-Z0-9_-]*\)\(\.ttf\),\2,g'); do
    subset_font $i 'woff'
    subset_font $i 'woff2'

Note the value of the --unicodes flag in the subset_font function. I had to figure out the Unicode code points I had to reference. This website on Unicode tables proved an invaluable resource.

About the two font families I use

This section is out-of-date. I now only use one font: Clear Sans. See commit 5f168cf6 for the details. However, FiraGO and Hack (alt) are still the default typefaces on my GNU/Linux computer. Refer to my dotfiles for more on that.

As I linked to my dotfiles already, I might as well write a few words about the fonts I have chosen.

The first is FiraGO, a sans-serif typeface, which is what is applied to the body text and headings. It is the main font of this website and is available in regular and bold weights with corresponding italics. This is also the system font I apply on my Debian machines running my custom working environment.

FiraGO is the continuation of Fira Sans, with support for more scripts and, possibly in the future, more variants. At this point, FiraGO has not yet deviated substantially from its predecessor, though this is to be expected.

Fira Sans was a project funded by Mozilla to cover the needs of the Firefox OS endeavour. It is a free/libre implementation of Erik Spiekermann’s “modern classic”: the FF Meta. Consider reading Matthew Butterick’s review on the matter.

My second font is Hack, a monospaced design that is ideal for long coding sessions because it is not flamboyant. It only tries to be utilitarian. Legible, clear, well balanced, decipherable. A true workhorse. It seems to me that Hack is the best typeface to emerge from the libre software milieu, as it is derived from the DejaVu fonts, which themselves trace their roots to Bitstream Vera.

The Hack designers offer the alt-hack repository for those who, like me, want to build a modified version of the font with some alternate glyphs. This is what I use (and what I distribute with my dotfiles, under the same license terms as the original). I have built the font from source and applied these patches:

  1. u0028-curved
  2. u0029-curved
  3. u0030-forwardslash
  4. u0033-flattop

To my eyes, these minor tweaks make Hack “sturdier” and are slightly more consistent with the overall personality of the typeface. This is just a matter of agreeing on the defaults. The original glyphs for those four code points add a bit more “character” to an otherwise Spartan presentation; the kind of flair I would rather avoid. Not that they are bad per se—just that I prefer their alternatives in the context of what Hack represents.

The right tool for the job

I will be refining my subsetting script and may build on top of it to automate various tasks, such as building the optimised woff and woff2 files for my website when new upstream versions are released.

Now that I have found a way to manipulate practically every available typeface out there, I can think of no good reason to ever compromise again with something like Google Fonts.

Granted, there is a bit of an effort involved, which is well worth it for the added flexibility it offers.