Speech Translation in human-to-human Interaction: Skype Translator

The error rate of speech recognition has come down enough, and the quality of machine translation has gone up enough to allow us to combine the two and achieve useful results. When building a translation system between humans who are not trained to talk via an interpreter, we need to take into consideration much more than the raw quality scores. Humans do not talk like machines, they do not even talk like the other humans who write down what they have to say: Spoken language is surprisingly different from written language. That’s a hurdle we can overcome in speech translation, by a technique we call TrueText: making spoken language look more like written language. Translates much better. Humans need an interface to work with and to interact with. There are multiple ways to address the spoken translation user interface, and many of them do not have the hoped-for result. How far can we expect the human to lean into the presence of machine interpretation? It turns out, not that far. We can design the speech translation system to be lenient and useful at the same time. With Skype Translator we have explored many aspects of user interaction and behavior, and arrived at two implementations targeting long distance, human-to-human conversations. Each implementation has advantages and disadvantages – we’ll discuss the two in comparison.