As of 2019, only 1% of the Internet is accessible in braille. Finding books, news articles and scientific papers in braille is a daunting task that millions of people around the world face every single day.
The reason for this is that braille is not a widely known language which only a small number of experts know well enough to make accurate translations. This makes braille translations extremely expensive and time-consuming to produce.
With most natural languages the translation problem boils down to picking the right words in the target language that will convey the meaning of the original sentence.
For braille, the problem is completely different. In braille, there are strict rules on how each word should be transcribed from the original language into braille. Some of these rules are relatively straightforward and rule-based software has been built that uses these rules to make simple translations.
However, there are more complex rules such as the rule that the braille translation should preserve the sound of the original word. Modeling sound is extremely difficult to do with a rule-based engine since the pronunciation of each word needs to be manually entered by a human. Furthermore, these pronunciation tables need to be constantly updated with new words that are being introduced into English every single day. This rule is very difficult to follow even by expert human translators as new words appear constantly.
Research into how to apply machine learning to this problem is very limited with no publically available datasets.
Since this is a unique problem that has received very little research attention we had to come up with our own algorithm to solve it.
Our first step was to build our own English-braille training corpus. We used over 600k sentence pairs translated by both humans and open-source rule-based engines.
We’ve then built an RNN-based neural network with an architecture specifically designed to model the sounding of words and to learn the rules of Grade 2 braille from the examples provided without any additional human input.
The final system is one of the best machine translators of braille to date, surpassing rule-based engines in the quality and speed of its translations. The system is capable of translating entire books within a few seconds and has been featured in the 2019 Cannes Festival.
To date the braille translator has received the following awards:
- Bronze Lions Award from the Cannes Festival 2019
- reddot "Best Of The Best Award" for Online/Digital Innovation 2019