Artificial intelligence converts typed text into handwriting

Scientists at UCL Computer Science in England have developed software that semi-automatically analyzes a piece of handwriting. The software ‘My Text in Your Handwriting’ can then convert typed text into the appropriate handwriting.

The machine learning algorithm to make sure that the ‘handwriting’ from the computer actually looks like handwritten text from a real person, analyzes glyphs or characters. Everyone makes letters in a different way and everyone has a recognizable style that is reflected in how the characters look and how the distances relate to each other.

The software not only ensures comparable letters, but also correctly formed ligatures or letters written together, dots, strikethroughs, the position of the letters on the writing line and the height of letters. The system also takes care of the constant small deviations, making the result almost indistinguishable from the real thing. The system does not use predefined fonts.

Participants in a blind test had to indicate whether it was a handwritten or a computer-written envelope. The 170 participants failed to identify 40 percent of the computer envelopes as counterfeit.

Lead author Gabriel Brostow tells the university’s news section that we don’t have to worry about forgeries, because this software actually gains additional knowledge about recognizing handwriting forgeries. According to him, the majority of handwriting recognition in forensic services is still done by hand. The algorithm can contribute well to this, he thinks.

The system needs ‘just a paragraph’ to learn a handwriting. The researchers wrote texts with manuscripts by Abraham Lincoln, Frida Kahlo and Arthur Conan Doyle, among others. Among other things, they had the text “Elementary, my dear Watson” attributed to Sherlock Holmes written.

The research article can be found in the ACM Transactions on Graphics journal.