Original URL: https://www.theregister.com/2010/02/16/darpa_echelon_request/

Echelon computers can't cope with bad lines

'What? Osama bin who? Yes, we sell nukes. Say again?'

By Lewis Page

Posted in Legal, 16th February 2010 14:08 GMT

Who remembers Echelon, the top-secret telecommunications spy network said to be run by the US and allied Anglophone nations, and to be triggered as soon as certain key words or phrases are spoken on the phone?

A lot of you, we'd guess. So it's interesting to note that Pentagon boffins have now stated that perhaps the most intriguing reputed capability of Echelon - the ability to automatically pick out words of interest and flag that conversation up as important to its human masters - doesn't work. Or anyway, it only works on good, clear lines: a noisy or degraded signal frustrates it.

The news comes as part of a solicitation from the Pentagon crazytech bureau, DARPA, in which the maverick military mayhem mavens request assistance with building a Robust Automatic Transcription of Speech (RATS) system. According to DARPA:

Existing transcription and translation and speech signal processing technologies are insufficient for working with noisy or degraded speech signals that are of importance to current and future Department of Defense (DoD) operations. Currently, there is no technological solution [our emphasis] which effectively addresses this kind of noisy and distorted speech signal, so operational units are forced to allocate significant human resources for this task.

One should note that America's feared National Security Agency (NSA, generally thought to be in charge of Echelon) is actually an arm of the DoD, not a civilian organisation.

DARPA says that the proposed RATS system should be able to tackle noisy audio signals and tell on its own whether they are speech or something else such as music. It should then be able to identify the language being spoken, and tell whether the speaker is a person of interest using voiceprint technology. Finally, the RATS software should be able to "identify specific words or phrases from a list of items in the language being spoken" - just what Echelon is supposed to be able to do already, only DARPA assure us that no such tech exists. Or anyway, none able to tackle a noisy signal.

So it would seem, if the DARPA announcement is on the level, that all you need do to evade the toils of Echelon - apart from making sure your call can't be intercepted, which is liable to be tricky - is introduce a bit of noise. At least until the RATS tech is developed, anyway, though as this is a "DARPA hard" project - ie so hard it's rather unlikely to be feasible - this will probably never happen.

Alternatively the DARPA announcement is merely a smokescreen to obscure the existence and capabilities of Echelon; or perhaps DARPA are merely looking to replicate those capabilities for use by less exalted military intelligence units in the field, as opposed to the mighty computer farms of the NSA in Maryland. The RATS requirement specifically mentions the languages Arabic, Farsi, Pashto, Dari and Urdu, suggesting that we might be talking here about conversations grabbed off satellite phones in southwest Asia, rather than ones gleaned from telecoms networks in the West.

However, as anyone knows who has used the voice-control/dictation tech which comes in smartphones and computers these days, the state of the art isn't terribly advanced for voiceware designed to run on normal platforms. The history of troubled voice-to-text firm SpinVox tends to illustrate this point too. DARPA may struggle to deliver a pocket, noise-proof Echelon to intercept teams working in and around Afghanistan.

And it's at least plausible that the real, big Echelon isn't actually quite as omniscient as people think.

The DARPA solicitation can be read here in pdf. ®