More on Skype Translator

Last month I blogged about Skype Translator. Since then, I had a chance to visit Microsoft in their Building 99 research facility to learn more about the technology behind the product. Next month I will talk more about its commercial potential.

Microsoft researcher Li Deng did a terrific job of walking us through the history and context for Skype Translator. Back in 1993, the general error rate for all machine translations of the spoken word was just below 100%, meaning that machines did a pretty poor job of translating the spoken word. By 2012, the error rate was down to around 7%, with progress accelerating in recent years:

Hitting the 7% threshold was significant enough that MIT Technology Review named the deep learning technology behind translator software as the #1 breakthrough development of 2013.

Dr. Deng believes that we are only a couple years away from reaching the limit for making "easy" gains in translator accuracy, with perhaps 5% being a commercially viable target. In other words, all the low-hanging fruit in translator improvement has been harvested, and further gains are going to take a lot of work. Another interesting observation is that the technology has hit the point where more data is not adding more accuracy. In my experience, this means that we are going to need more brilliance rather than more compute power in translator product design before we can get another significant advance in the technology. Companies like Microsoft and Google are collaborating on the science of deep learning, so as breakthroughs are made it is likely to show up simultaneously in their respective products.

Until that brilliance arrives, that probably means we need to build UC products that rely on translator technologies that can be effective and commercialize with a 5% error rate. I don't yet know how this will play out, but it has already been observed that translator software is probably good enough for casual conversation but not for professional translation (http://verbalink.com/services/translation-services#translation-infographic). This then raises the question of how good the translators need to be to work effectively enough in call centers and other circumstances where conversational accuracy of 95% is sufficient to get the job done.

My take on this is that 95% accuracy is good enough for anything short of business / legal negotiation (where craftsmanship of words has a very high standard) and in situations where grammatical errors and modest mis-translations will be tolerable. I know from experience that my Spanish is 60% awful, yet I manage to use it effectively enough for most vacation-related activities like booking reservations and arguing about bar tab bills. So there is a lot of upside here for UC products that embed translator technologies.

More on this in my next blog.

Also on UCStrategies.com on this topic:

• Musings on Skype Translator, by Joseph Williams
• The Business Case for Speech Translation in UC, by Joseph Williams