She speaks English, is heard in French, translated … Skype – The Obs

You say “hello”, she hears a synthesized voice that pronounces “hello”. And vice versa. We tested the first interview with simultaneous English / French translation via webcam. Today, Microsoft is launching this new feature on Skype, revolutionary and accessible to all (with a version of Windows 8) in beta.

We conducted this experiment with a Microsoft linguist Kara , originating in Canada, which was 8,000 kilometers away, in Seattle (we were at the headquarters of Microsoft France in Issy-les-Moulineaux). The aim was to hold a “banal” conversation. So we improvised

Watch the first interview with simultaneous translation translator Skype.

First impressions? Confusing and dizzying

Confusing because it must get used to the synthetic voice that translates everything in near real time. – One can also read the translation in writing, via the chat tool

Confusing also because the software still running is far from all understand exactly -. we suspected. For example, the software can “miss” some words spoken too quickly, pronounce sentences or straight came from another planet.

Dizzying as well because when translation proves perfect – and it works better than what you could imagine – it was like living a magic moment, historical

And you say it’s not that. a start. This is confirmed by the technical director of Microsoft France, Bernard Ourghanlian: “It was opened in beta (note: test) , since the machine still needs to learn. However, the software feeds conversations. We feel that we are missing 4,000 hours of discussions between an anglophone and a francophone to refine the tool

You can also call today. – And theoretically understand – a correspondent speaking Italian, Mandarin, German and Spanish!

Microsoft engineers imagine many practical applications. For teachers: children across the world can communicate and understand other cultures. Or to NGOs, or for those who rent their apartment on Airbnb. But also to communicate with deaf people without going through sign language. – Reading simultaneous translation in writing

The challenge is immense, says Bernard Ourghanlian Microsoft because the spoken language is different from the written. When we say “I do not know”, software tended to understand “I know” because it lacked the “ne”.

This may seem surprising, but the software easier to understand and translate long sentences because there is more context “. Indeed, when we asked our interlocutor what we eat in Canada, Skype hesitated, but when properly translated added “for example for lunch or dinner.”

One biggest problems for simultaneous translation: capturing an intelligible sentence. Eliminate the “ben”, “uh”, and ambient sounds, which disrupt the proper understanding of the sentence. The microphones equipping most computers, very basic, do not help to understand everything artificial neurons. “On the Xbox, where you can also use Skype, there are 4 good quality microphones, it changes everything,” Microsoft advance. The quality of bandwidth also comes into play

Finally, the machine must also detect the intonation if a question is asked, for example:. Voice synthesis must reproduce the nuance of your tone.

“With our new neural system of” deep learning, “we have gone from a 20% rate of errors 7%, does one slip at Microsoft, confident of progress come.

Aurélien Viers