Normally we assume that a call will be anonymous if you buy a prepaid cell phone in cash, and call someone who’s also bought a prepaid phone in cash. It turns out that isn’t true, at all. A Russian company has started selling voice recognition systems to anyone who’ll buy —
systems that can identify the speakers in an intercepted call with 90% accuracy. (and if it’s available to the private sector, guess what that means for other attacker classes?)
The obvious countermeasure, of course, is text-messaging. It’s a lot harder (but not impossible, given enough text) to fingerprint stuff like ‘R U OK?’, though intercepting and analyzing texts is far cheaper than voice. (This may be one place I’d consider it almost OK to use a smartphone — sending email from your own server could add location privacy that texts don’t allow.)
More useful, though, would be some good research into voice changers —
identifying how voice recognition systems identify a voice, and then figuring out how to distort it so that the speech is comprehensible but mechanically unclassifiable.
While Samuel Jay Keyser might be a starting point (in more ways than one…), Patrick Flanagan’s Neurophone (remember that project) could be another. By essentially transforming voice into a 50kHz NBFM signal, Flanagan discarded a tremendous amount of information while still keeping things intelligible. Indeed, he spends rather a lot of the 2nd gen Neurophone patent ruminating on the possibilities inherent in this — in his case, designing a delay-line-based voice encryption system.
The question is, could you do the same thing but keep the signal within the few-kilohertz bandwitdth of the phone system?
“a countrywide biometric database designed to store millions of people’s “voice-prints.”
Russia’s Speech Technology Center, which operates under the name SpeechPro in the United States, has invented what it calls “VoiceGrid Nation,” a system that uses advanced algorithms to match identities to voices. The idea is that it enables authorities to build up a huge database containing up to several million voices—of known criminals, persons of interest, or people on a watch list. Then, when authorities intercept a call and they’re not sure who is speaking, the recording is entered into the VoiceGrid and it comes up with a match. It takes just five seconds to scan through 10,000 voices, and so long as the recording is decent quality and more than 15 seconds in length, the accuracy, SpeechPro claims, is at least 90 percent.
The technology has already been deployed across Mexico, where it is being used by law enforcement to collect, store, and search hundreds of thousands of voice-prints. Alexey Khitrov, SpeechPro’s president, told me the company is working with a number of agencies in the United States at a state and federal level. He declined to reveal any names because of nondisclosure and confidentiality agreements. But Khitrov did divulge that various versions of the company’s biometric technology are used in more than 70 countries and that the Americas, Europe, and Asia are its key markets. Not all of its customers are law enforcement agencies, either. SpeechPro also designs voice recognition technology that can be used in call centers to verify the identities of customers. Depending on the size and specifics of the installation, it can cost from tens of thousands up to millions of dollars.[…]
Speech Technology Center’s products have been sold to countries including Kazakhstan, Belarus, Thailand, and Uzbekistan—hardly bastions of human rights and democracy. What if the VoiceGrid Nation system were in the hands of an authoritarian government? It has the technical capacity, for example, to store a voice-print of every single citizen in a country the size of Bahrain—with a population of 1.3 million—which would allow state security agencies to very effectively monitor and identify phone calls made by targeted political dissidents (or anyone else for that matter).”