This image shows how VoiceGrid works and here’s some other info gleaned via their documentation. Voice matching technology can “automatically separate the voices within a two-person dialog and send each voice individually for matching” and is being used as “part of a comprehensive plan to best leverage existing and new audio data.” Even without considering the NSA surveillance via intercepting calls, the whitepaper gives numerous examples of passive sources for voice recognition data that has “already been collected.” These include voicemail, recordings made while speaking to commercial service providers such as banks, cell phone companies, and cable TV companies, as well as 911 calls, suspect interviews and court recordings.
The company’s technology uses three methods for voice matching and an algorithm that automatically compares “voice models against voice recording obtained from different sources such as cell phones, land lines, covert recordings and recorded investigative interviews.” When combined, there is a 90% voice match to identification accuracy within 15 seconds. However, according to VoiceGrid’s “key figures,” it only takes: · 3 seconds is the minimum required speech pattern for analysis.
· In 5 seconds, it can search/match in 10,000 voice samples.
· 10 seconds is the average time for feature extraction.
· Executes up to 100 simultaneous searches.
· Accommodates up to 1,000 active users.
· Stores up to 2,000,000 samples.
(more)
