Innovative chip architecture allows fast, automatic voice recognition

Published 22 August 2006

Government security agencies are engaged in ever-more-expansive data mining operations; as more voice and video digital media goes online, there is a growing challenge of how to search these live streams effectively; the audio mining problem is especially acute in homeland security applications which need to expedite translations of huge streams of incoming data, sifting for a few critical words or phrases

Durham, North Carolina-based Semiconductor Research Corporation (SRC), a leading university-research consortium for semiconductors and related technologies, said that researchers at Carnegie Mellon University (CMU) have developed a chip architecture which allows automatic recognition of voice streams at a rate which is up to ten times faster than computers currently supporting speech recognition. The implications of this breakthrough are important for both the commercial chip industry and the defense and homeland security communities. The research has been targeted at creation of a radically new silicon chip architecture which is exclusively dedicated to speech recognition and performs it more efficiently than current state-of-art computers.

Professor Rob Rutenbar, who leads the In Silico Vox project at CMU, explains: “Whether running an enterprise-class voice call-in center or decoding individual words on a cell phone, all of today’s serious speech recognizers exist as software running on some processor. That’s terribly limiting…. Moving these computations directly into silicon means we can perform recognition dramatically faster, cheaper, and better for both commercial and homeland security tasks.” Dr. David Yeh, director of Integrated Circuits and Systems Sciences at SRC-GRC and a Texas Instruments assignee to the consortium, agrees: “Speech recognition has a history of struggling to meet expectations but this research enables dual-purpose applications that were previously untouchable in the consumer space and for purposes of combating security threats to the country.”

We reported about the growing interest of government security agencies in data mining. As more voice and video digital media goes online, technology users face the increasing challenge of how to search these live streams effectively. This increasingly daunting audio mining problem is especially acute in homeland security applications which need to expedite translations of huge streams of incoming data, sifting for a few critical words or phrases. The fast-recognition technology being perfected at CMU is aimed at making it possible to rapidly search massive volumes of data. Many now working at the NSA, CIA, DHS, FBI, and the Pentagon would be garetful.

-read more in this news release