»language«

Looks like a prototype model for James Bond: the glasses with a small display added show the translated speech of a person talking to you in a foreign language. It goes as follows: the speech is recorded by a microphone [1], transferred into text [2], translated [3] and given as subtitles on the display [4]. The key component of course is the software that performs the speech-to-speech translation [step 3] as speech recognition and output (or synthesis for an ear plug) already work reasonable good. The prototype is developed at U of Karlsruhe, Germany, in the group of Alex Waibel. He told me recently that the performance in the language pair english/spanish is already fine, for instance, to follow a lecture. The pair english/german is moderate just to grasp an idea on what the speaker is saying.

Interesting is the software's approach to 'learn' a foreign language: just assume that any sentence or phrase is already translated somewhere on the Internet. Now, let the search engines collect the paired phrases and do some intelligent data interpretation. If it works for one language pair, it will do it for any other pair with only slight modifications of the software, hopes Waibel. image (c) University of Karlsruhe