It would normally pass unnoticed that researchers here in Pittsburgh will chat over the Internet today with colleagues in five other countries. They'll be talking about travel arrangements. Boring stuff, really, except for one crucial difference:
Everybody will be speaking their own language but will hear everyone else translated automatically into their native tongue: English to German, Korean, and so on.
For decades, scientists and science-fiction writers have dreamed about a translating machine that would breach the world's language barriers.
But the dream has proved frustratingly elusive. As every junior-high French student knows, translation involves many more skills than simply plugging in foreign words for English ones. The smartest computers have yet to master any one of those skills, much less put them together with any sense of fluency.
But two compensating strategies are allowing researchers to build useful approximations of the universal translator. Today's demonstration here at Carnegie Mellon University and five other research institutions in France, Germany, Italy, Japan, and South Korea illustrates both techniques.
In a large room on the Carnegie Mellon campus, graduate student Chad Langley talks into a microphone: "Hello, my name is Chad. I'd like to go to Heidelberg, please."
Barely an instant later, on the other side of a room divider, project scientist Monika Woszczyna hears a German translation. She answers in German and on his side of the divider, Mr. Langley hears, in English: "When would you like to go?"
Computer plays travel agent
Back and forth, the German-English conversation flows. Langley, role-playing a would-be traveler, wants to leave July 15 and return the 22nd. He books a flight and a hotel room. He wants tourist information on the local castle. Ms. Woszczyna, acting as travel agent, responds each time in German.
The small demonstration, held last month on the Carnegie Mellon campus, shows how far researchers have come. Their system now handles 10,000 words, double its capacity a few years ago. Its responses are faster and more accurate. Demonstrators don't have to script their speech to get satisfactory results.
Still, it makes mistakes, saying at one point that the plane will "flight" Langley at 5 p.m. More important, it handles only a single subject: travel-agent requests. Ask it if humans will ever fly to Mars, and the machine would not know how to respond. By restricting translation to a single topic or domain, researchers are starting to create systems good enough to leave the laboratory and enter the working world.
Their other compensating strategy involves visual clues to help the listener when the speech isn't quite right. For today's international demonstration, the Carnegie Mellon group has attached Web-looking documents on a screen that its partners will see. The Pittsburgh-based researchers have even included a little window that displays the speakers' facial expressions using very little bandwidth.
"It's more than a proof of concept," Woszczyna says of the demonstration. "We really want to show ... how all these things work together."
Refugees talk through laptop
Universities are not the only entities working on speech-to-speech translation.
When peacekeepers moved into Bosnia in the mid-1990s, Western doctors had to question patients, and military officials had to interrogate locals. But there weren't enough translators to go around. So in 1996, the Defense Advanced Research Projects Agency came to Dragon Systems, a leading speech-
recognition company in Newton, Mass., for help. The company sold its products in several languages, but it hadn't tackled translation.
Nevertheless, over a weekend, the company had cobbled together a crude but workable system to demonstrate. Soon the laptop-based systems were shipped to Bosnia, first for doctors, later for land mine-clearing units. As universal translators, they hardly measured up. The translation worked only one way: English to Serbo-Croatian. And the foreign-language output was rudimentary: some 4,000 to 5,000 stock phrases such as "Do you speak English?" "Where does it hurt?" "Do you know of any land mines around here?"
Locals would have to point or gesture in response.
"Things like that were very effective because you can get a long way without speech," says David Wald, senior research engineer with Dragon. The company is waiting to ship over similar English-Albanian translation systems for peacekeepers in Kosovo.
Another speech-recognition company, Belgium-based Lernout & Hauspie, is hard at work on several translation projects. It sells a general program, called L&H Power Translator, that gives a rough-draft approximation of documents in English and five other languages. But for better accuracy, the company also offers industry-specific software with specialized vocabularies.
"It's not perfect" - and it's text-based, not speech-enabled, concedes Gaston Bastiaens, the company's chief executive officer.
"But as long as you remain limited in the domains, you can do pretty good communication.... We think five years from now we will be so far along that you will speak in English but someone in Beijing will hear you speaking Mandarin."
But a universal translator with the ability to translate any subject from tennis to Tennyson? "Once you start discussing poetry, then I think we will disappoint you," Mr. Bastiaens says. "We are many years from that."
(c) Copyright 1999. The Christian Science Publishing Society