Military could be using high-tech speech software by 2017
WASHINGTON — The Pentagon could be able to listen in on voice communications in difficult environments and then quickly translate and transcribe them for use by intelligence analysts and combat troops by 2017, according to the Defense Advanced Research Projects Agency.
Newly released DARPA documents show it is continuing the next two stages of its Robust Automatic Transcription of Speech program, which is aimed at separating speech from background noise, determining which language is being spoken and then isolating key words from that speech for analysis.
The Air Force, DARPA says, is testing the third phase of the program in the field now, while "the research division of a government agency will be testing the speech activity detection algorithm to incorporate into their platform." References to "a government agency" usually refer to a part of the intelligence community, such as the CIA or National Security Agency.
So far, DARPA has spent $13 million on RATS. It now wants to spend another $2.4 million, contract records show, to make the final push to make the system operational by the Air Force as early as this year but by 2017. Other agencies, particularly those in the intelligence community, will also use the system once it's operational.
The fourth and fifth phases, DARPA says, will take the system into the field to determine the difference between real speech picked up in the field and that developed for the laboratory. The latest developments, the agency says, work well in the field but suffer "a great degradation in accuracy from the performance on lab data."
The program involves four key areas, DARPA documents show:
• Speech activity detection. What is speech and what is background noise?
• Key word spotting. Analysts can be directed to seek out certain words, such as in the 2007 movie The Bourne Ultimatum, in which NSA computers tracked the word "Blackbriar," which identified a secret government program.
• Language identification. Once the software has separated speech from noise, what language is being spoken? DARPA researchers are focusing on Arabic, Farsi, Pashto, Dari and Pashto, which are used in the Middle East, Afghanistan and Iran.
• Speaker identification. The system can use voice-pattern analysis and other technology to determine who is speaking, which can be particularly useful if intelligence analysts and troops are looking for certain targets.
Much of the work on the RATS system has been done in laboratories in quiet and controlled environments. Problems, DARPA documents show, arise when researchers try to collect voice signals in environments with a lot of background noise and competing radio signals.
Troops and intelligence analysts trying to monitor multiple signals need to be able to separate one signal from another and then focus on the actual words being spoken without the interference of static or background noise. That's challenging in controlled environments but even more so in places like Afghanistan or Iraq.
Once the technology can clean out the background noise and static, it must rapidly translate the spoken words, which are often in regional dialects and accents that are hard to understand for outsiders.
DARPA has pushed voice-recognition research since 1971. Much of the technology paid for by military research funds have spawned developments such as the voice-recognition software used by the Apple iPhone's Siri system. It's getting us closer to the kind of universal translator familiar to viewers of Star Trek for almost 50 years.
The RATS program involves an international coalition of contractors and scientists led by BBN Research, a division of defense giant Raytheon. It involves the Czech Republic's Brno University of Technology, Johns Hopkins University in Baltimore, the University of Maryland and Carnegie Mellon University in Pittsburgh.
Other companies and universities involved in the project include IBM, SRI, the Massachusetts Institute of Technology and universities in Denmark, England and Spain.