voice mining

Speech technology is revealed to have a deeper role in the warrantless wiretaps (from NYT, emphasis mine):

Last September, the N.S.A. was granted a patent for a technique that could be used to determine the physical location of an Internet address — another potential category of data to be mined. The technique, which exploits the tiny time delays in the transmission of Internet data, suggests the agency's interest in sophisticated surveillance tasks like trying to determine where a message sent from an Internet address in a cybercafe might have originated.

An earlier N.S.A. patent, in 1999, focused on a software solution for generating a list of topics from computer-generated text. Such a capacity hints at the ability to extract the content of telephone conversations automatically. That might permit the agency to mine millions of phone conversations and then select a handful for human inspection.

As the N.S.A. visit to the Silicon Valley venture capitalists this month indicates, the actual development of such technologies often comes from private companies.

In 2003, Virage, a Silicon Valley company, began supplying a voice transcription product that recognized and logged the text of television programming for government and commercial customers. Under perfect conditions, the system could be 95 percent accurate in capturing spoken text. (NB: This is almost unheard of; 10% transcription of such degraded audio would be more normal) Such technology has potential applications in monitoring phone conversations as well.

Español | Deutsche | Français | Italiano | Português| Ch| Jp| Ko