Arstecnica
Thanks to a free web app called calligrapher.ai, anyone can simulate handwriting using a neural network running in a browser via JavaScript. You type a sentence and the site renders it as handwriting in nine different styles. Each style can be adjusted with properties such as speed, readability, and stroke width. You can also download the resulting fake handwriting sample as an SVG vector file.
This demo is particularly interesting because it doesn’t use fonts. Typefaces that look like handwriting have been around for more than 80 years, but no matter how many times you use them, each letter is repeated.
Over the past decade, computer scientists have eased these limitations by discovering new ways to simulate the dynamic diversity of human handwriting using neural networks.
Created by machine learning researcher Sean Vasquez, the Calligrapher.ai website draws on research from a 2013 paper by Alex Graves of DeepMind. Vasquez created his Calligrapher site a few years ago, but it recently made headlines when it was rediscovered on Hacker News.
-
An example of handwriting synthesis from the Calligrapher.ai website.
Calligrapher.ai
-
An example of handwriting synthesis from the Calligrapher.ai website using different styles.
Calligrapher.ai
-
The handwriting on this computer is terrible because it’s becoming less legible.
Calligrapher.ai
-
Improved readability makes the characters clearer.
Calligrapher.ai
Calligrapher.ai “renders” each letter as if it were written by a human hand, based on statistical weights. These weights come from a recurrent neural network (RNN) trained on the IAM online handwriting database. The database contains 221 handwriting samples of his, digitized over time from whiteboards. As a result, Calligrapher.ai’s handwriting synthesis model is heavily tuned for English writing, and the folks at Hacker News report having trouble reproducing diacritics commonly found in other languages. doing.
The algorithms that generate handwriting are statistical in nature, so properties such as “legibility” can be dynamically adjusted. Vasquez explained how the readability slider works in his Hacker News comment in 2020. A common technique is called “adjusting the temperature of the sampling distribution”. ”
With neural networks tackling text, speech, photos, videos, and handwriting, every inch of human creative output doesn’t seem beyond the reach of generative AI.
In 2018, Vasquez contributed the underlying code that powers the web app demo on GitHub, so it could be adapted for other applications. In the right context, it can be useful for graphic designers who need a better flair than static script fonts.
