Skype’s Real-Time Translator Learns How to Speak From Social Media

Deciphering social media slang can be challenging, especially when attempting to translate it. That’s the issue tackled by Microsoft researchers, who have discovered a unique method for teaching the forthcoming Skype Translator to communicate more like humans by drawing upon social media information.

The researchers argue that social media could be instrumental in enhancing computer understanding of human communication. Work in the sphere of computational social science has shown that it’s potentially viable to extract social meaning from speech and text in complex natural tasks, as evidenced by social media experiments. Applications like the Skype Translator, slated for beta release soon, promise real-time translation of multilingual conversations.

However, getting it right proved to be challenging. Past approaches to bridge speech recognition, machine translation, and speech synthesis proved unsuccessful due to the fundamental differences between spoken and written language. Microsoft Research therefore sought out innovative techniques to close this gap, one of which was a software system designed to translate social media content.

This system, phrasal statistical machine translation (phrasal SMT), would learn from n-gram alignment and then spit out the most probable translation when presented with new, untranslated text. However, it was not without flaws, the most significant being confusion when the words in an uncommon phrase were reordered. This led the Microsoft Research team to develop a system called syntactically informed phrasal statistical machine translation (syntactic SMT), which brings in grammatical understanding.

Even with this improvement, challenges remained, particularly when it came to translating communications from social media platforms like Facebook, SMS, and Twitter, each with their unique characteristics. The team had to develop a text normalization system specific to social media, which could adapt to these style variations and produce content that the syntactic SMT could process. This technology increased the accuracy of social-media text translation by 6 percent.

Microsoft’s Skype Translator isn’t alone in the landscape of speech translation systems. Google Translate, for instance, is also used by many to test their command of foreign languages. However, unlike Skype Translator, Google’s system doesn’t yet offer real-time translation integrated into a video telephony application. Only a handful of companies worldwide have the necessary resources to develop and scale a system like this, capable of serving millions of users.

Updated in 2025 to align with recent developments.