Slashdot Mirror
Can Your Mouth Become Multilingual?
Roland Piquepaille writes "During a videoconference last week between Karlsruhe, Germany, and Carnegie Mellon University (CMU), Pittsburgh, USA, the talk of Alex Waibel, from CMU, was automatically translated in German and Spanish. Both the Pittsburgh Post-Gazette (PPG) and the Pittsburgh Tribune-Review (PTR) attended the conference, took pictures and were impressed by this new 'open domain' speech-to-speech translation. This new computer technology is based on artificial intelligence (AI) and statistical methods. During the demonstration, the speaker had electrodes attached to his face and his neck, but the researchers think that these electrodes could be implanted into your mouth and your throat in a decade from now -- if you agree of course."
Computers have a hard time translating written things as it is... any bilingual will tell you that online translators for complete sentences will do nobody any good, for the most part. My Spanish teachers are all able to see papers with computer translations very easily, due to similarities in words and meanings (such as the word "pants" which can be colthing or breathing heavily) Not to mention, grammar and things like that are not done well at all. For the fun of it, try going to an Online translator and write something in English, translate it to Spanish, then back to English. Some results are pretty crazy. I guess the point I'm trying to make is this: what makes the translators so special compared to the ones we have now? How can they work better? Sure, there is probably a bit more effort put into these, but I don't think that a good translator will be available for another 5 years, not to mention the whole "take the speech you aren't saying" thing is hard to believe.
Another approach is from some work I saw demoed at an MIT conference in Vienna. If you capture enough video of a person speaking, you can remix/rerender video of that person saying anything you want them to say. The software works at the phonetic level so you can even synthesize words that the person has never even uttered before and even make them appear to speak languages that they don't know. They had some visually convincing video showing people saying things that the researchers claimed they never said. Yes, the demo version worked with clean test video and a professional video/image analyst could probably spot a faked/remized video. But if these technology becomes good enough, I can see it making video a nontrustworthy source of data (like skillfully retouched photos).
Two wrongs don't make a right, but three lefts do.
Hey,
...). Such a combination would be suitable both to deal with large amounts of data (statistical) and to deal with negation (only available in the symbolic appoach).
Alex Waibel was one of the leading scientists in the Verbmobil project in 1995. The technology was pretty interesting (maintaining probability "graphs" from the Markov speech analysis through the syntactic and semantic analysis).
However, results were pretty poor due to the structure of the project (just too many people) and because many institutions really weren't interested in the project and went for their favourite research topic with a new name (that's how research in Germany works...). Perfectly possible that Mr. Waibel advanced with the topic, now 10 years after the first major trial...
Personally, I actually gave up AI completely after the ESSLLI (European Summer School on Logic, Language and Information) and promised not to touch the subject again until there were a "unified" formalism incorporating the old "symbolic" approach (predicate logic etc.) and the new statistical methods (Bayes, Markov,
Maybe they've got it this time? It's a pitty they don't talk more about the underlying formalism.
Btw., the electrodes are probably just an enhancement of the normal speech recognition software to get a better "signal".
Bests,
Frank
http://www.project-open.com/