Letting machine learning choose the right font for everyone

Set of ampersands.

Image credit: Adobe Stock/ PureSolution.

Did you know you have the ability to read 20 percent faster just by changing the reading font? Which fonts are effective for reading depends largely on the individual, and while age is a significant factor, so is reading proficiency (speed, frequency, etc.), familiarity with a particular font and style, and other reader characteristics like reading difficulties.

Given all the factors involved, and more than 800,000 digital fonts available to choose from, it can be difficult and overwhelming to know which font is optimized for your reading style. But what if a machine learning (ML) algorithm could handle the decision making for you?

In Personalized Font Recommendations: Combining ML and Typographic Guidelines to Optimize Readability, recently published in the ACM DIS’22 journal, we conducted a study to help inform which font characteristics work particularly well for different types of readers. The findings helped to build these relationships into a model to recommend faster fonts for specific readers. The study was a collaboration between vision scientists, data scientists, and typographers. “It’s really exciting to explore whether machine learning could be harnessed to personalize reading experiences,” said lead author and Acrobat.com ML engineer, Tianyuan Cai.

Using machine learning to personalize reading

We interviewed several typographers to understand which design features make specific fonts particularly optimized for reading. Based on our interviews, eight fonts were selected, and 252 remote participants (ages ranging from 18 to 71) read passages and answered comprehension questions.

Examples of SERIF font and SANS SERIF.

Example of 8 fonts selected for a study.

Above are eight fonts selected, in consultation with typographers, for our study, which focused on digital reading. Four serif fonts approximately match the look and feel of the four sans serif fonts, including fonts that have perceptually different sizes. Typographers control the appearance of fonts by adjusting font metrics. In the second image we see how researchers visualize the font metrics under consideration for our study and model.

Like findings from previous research, participants struggled to identify a faster reading font by preference alone. In our study, participants’ most preferred fonts are, on average, just as fast as a font selected at random. This result strengthened our motivation to find an automatic way of recommending fonts to readers. Our ML model predicts fonts that help people read approximately 23 WPM faster than a randomly chosen font, and 159 WPM faster than their slowest font.

These are real gains that our model achieves by learning to relate reader characteristics to font characteristics:

In this work, we relate these characteristics by comparing the eight fonts tested and support our findings with prior literature. More work is underway to evaluate the model with a larger set of fonts and testing for repeatability of these findings.

Model that uses characteristics of readers and fonts to predict suitable fonts for individual readers.

Our model uses characteristics of readers and fonts to predict suitable fonts for individual readers. Age is a factor that the model considers relevant, but also the readers’ self-reported reading speed, frequency of reading, and familiarity with the font in question.

Will users actually be open to the idea of choosing their reading font with the help of machine learning? 90 percent of our study participants agree that personalizing fonts can lead to improvements in reading, and 86 percent said they would trust a font recommendation if it was generated by a computer algorithm. Thus, the way is clear for the text personalization algorithms of the future, and it is up to us to help build them.