Darpa Looks to Build Real-Life C3P0

(Wired) Right now, troops trying to listen in on enemy chatter rely on a convoluted process. They tune into insurgency radio frequencies, then hand the radio over to local interpreters, who translate the dialogues.  It’s a sloppy process, prone to garbled words and missed phrases.

What troops really need is a machine that can pick out voices from the noise, understand and translate all kinds of different languages, and then identify the voice from a hit list of “wanted speakers.” In other words, a real-life version of Star Wars protocol droid C3PO, fluent “in over 6 million forms of communication.”

Now, the Pentagon’s trying to fast-track a solution that could be a kind of proto-proto-prototype to our favorite gold fussbudget: a translation machine with 98 percent accuracy in 20 different languages.

Darpa, the military’s experimental research agency, is launching the Robust Automatic Translation of Speech program to streamline the translation process. (That’s “RATS,” for short. Ouch.)

It’s the latest in a series of Darpa-funded ventures that want to turn foreign-language text and speech into helpful data for English-speaking GIs, and vice versa. In the early days of the Iraq war, Darpa’s handheld Phraselator gadgets allowed simple English phrases to be spit out in Arabic. These days, they’re testing prototypes of iPod-sized two-way, speech-to-speech translators, for easier communication between troops and foreign civilians. The gadgets are 70 percent accurate at translating Iraqi Arabic into English, but depend on a clear signal and little to no background noise.

In 2008, Darpa awarded $5.6 million to BBN Technologies for Madcat, an automated system that works on a laptop or PDA device to do quick text translations — anything from sidewalk graffiti to scribbled notes — in a war zone. Darpa’s made quick progress with Madcat, and they want RATS to accomplish equally impressive feats with speech. Much like Madcat can translate even the sloppiest, most illegible writing, Darpa wants RATS to pull speech out of “noisy or degraded signals,” with 99 percent accuracy at distinguishing spoken words from background noise. The system should also be able to identify, with 98 percent accuracy, the language spoken — with special emphasis on Arabic, Farsi, Pashto, Dari and Urdu.

 The RATS software will be programmed with voice-recognition technology, to identify people on a military most-wanted list. It’ll also be able to automatically detect specific, preselected “key words or phases.”

Darpa’s looking for prototypes that live up to some lofty expectations: Within six months, they want demos that translate 15 languages among 1,000 different speakers, and identify 100 words or phrases in Arabic, Pashto and Farsi.

The software would have tremendous application in military circles, and trim the Pentagon’s billion-dollar budget for human translators. But it might be coming sooner to civilians, thanks to Google. The company’s already mastered multilingual website translation, and last November they ventured into voice recognition with 1-800-Goog-411. The iPhone app allows users to “search” using voice commands on their cellphone. Each search “teaches” Google’s computers, improving their voice-recognition abilities.

Right now, Google’s app only works in English, but they anticipate a multilingual, cellphone translation system within a few years. And Darpa’s only budgeting 18 months for the completion of RATS. With two of the world’s biggest tech-agency superpowers throwing millions into the idea, civilians and war-fighters alike might soon be speaking a universalized language. Can R2D2 be far behind?