DARPA Awards Contract to Develop Robust Speech Processing Software for Noisy Environments
SRI International has been awarded a $13M contract from the Defense Advanced Research Projects Agency (DARPA) to develop software to process noisy and highly degraded speech. The project was awarded as part of DARPA’s new Robust Automatic Transcription of Speech (RATS) human language technology research and development program.
For this project, SRI’s Speech Technology and Research (STAR) Laboratory will develop speech processing algorithms and software to detect speech activity, identify the speaker and the language being spoken (from a predefined set of speakers and languages), and search for key words. The goal of the program is to extract information from highly distorted audio signals, in a variety of languages. These types of audio signals may be found in many military and civilian contexts, such as air traffic control, emergency and ship radio communications, and call-in segments of talk shows.
To tackle this technology challenge, SRI has formed a team including leading academic speech research groups from Carnegie Mellon University (CMU), the Berkeley-based International Computer Science Institute (ICSI), the University of California at Los Angeles (UCLA), and the University of Texas at Dallas (UTDallas).
"Noisy environments and degraded transmission channels pose an enormous challenge to speech processing technology, which means that government and military personnel have to dedicate a significant amount of time analyzing distorted audio recordings," said Horacio Franco, chief scientist at SRI. "The difficulty is exacerbated by the need to handle a mix of languages, some of which, like Dari and Pashto, are not covered by existing commercial software and speech databases. The software that SRI and its partners are developing will be beneficial for defense-related operations that need to gather information and intelligence, such as speech recognition or voice-based speaker verification from a cell phone on a noisy street corner."
