Untrained speech recognition on phone quality voice is far too crappy for pulling some word like “murder” out of a conversation. All the masons at construction sites talking about “mortar” would constantly be seeing black helicopters descending silently towards them on autorotation.
I’d have thought that you use the speech recognition to trigger the recording for a MkI bum on a seat to evaluate and feed back into the speech recognition to improve the hit rate.
Sure, but even using the automation to trigger human intervention I still don’t think it’s anywhere close to feasible. Several orders of magnitude off, IMO. And needless to say, anyone with a grain of sense (admittedly this rules out many terrorists) would use code phrases anyway, so this would be a vast expenditure for a net to catch only the stupidest of fish.
Technically speaking it is very feasible. A single server not much more powerful than a high end gaming PC and equipped with the right DSP cards can real-time process voice recognition on hundreds of simultaneous audio channels. It wouldn’t be that hard to put a reasonable filtering algorithm in place so that even with all the false-positives from computer speech recognition only the most suspicious calls got flagged for analysis by human operators.
I will agree that it’s probably inadvisable and wouldn’t catch anyone smart enough to use code words anyway, but the telecom tech nerd in me wanted to break down how such blanket monitoring could be implemented :)
I meant it’s not feasible by cost and effectiveness. They can always throw hardware and even people at the problem within their budgets, but that just won’t accomplish anything useful. I worked with Nuance servers at Verizon Labs, myself… Sure you can have a grammar sitting there waiting to fire while listening for a list of a hundred keywords or whatever you like, but 99.999% of the conversations that mention murder or bombing or whatever will be from ordinary civilians, and the remaining 0.001% will be criminals so stupid they are probably already being monitored by live personnel anyway. So it just makes no sense to implement the program in the first place.
Oh sure, I don’t think we disagree on that end of it. It would be a ludicrously ineffective use of money in terms of the results, but so it a lot of other silly stuff the government decided to do in a post 9/11 panic. It wouldn’t surprise me too much to know that the NSA was actually doing something along these lines.
More likely the keyword recognition is used as an adjunct to other intelligence gathering activities, like finding out which phones are being used by people uttering a particular word or combination of words or phrases that has previously been used by a known terrorist suspect or organization. Coupled with other data-mining results this could be effective.
I also wonder if voice-recognition software can identify a particular person’s voice, no matter what words they are saying. This could be handy if you were searching for someone who might be using a prepaid or disposable cellphone and you wanted to track their movements.
Conventional “voiceprints” are not very useful, and I bet they’re even worse when they have to deal with crappy phone voice with some bands chopped off (though do they still chop up the signal on modern mobile networks – maybe not).
When I was at Bellcore 20 years ago, I heard a research claim at a talk that there was a sort of signal underlying speech sounds that would make for a much higher quality voice-based identifier, but I don’t think the research ever went anywhere useful.
Indeed, “ludicrously ineffective use of money” is something of a Washington mantra. It was twenty some odd years ago when I was working with intel types, and I’m sure it hasn’t gotten any better today. So just because something is financially insupportable doesn’t mean it ain’t being done.
Wait, so was Project Echelon all run with human beings listening to handpicked important people? Maybe in the beginning, but by the 80s they must have automated some processes?
I don’t think computers good enough to do any realtime speech recognition were available in the 80’s. I’d say 90’s would have been the first possible automation.
I’m not sure there is public knowledge of the true story about Echelon, especially whether it ever did anything worthwhile, but it’s infinitely easier to pick words out of email and telex than it is to pick words out of conversation. So far as I know it was never applied to voice.
There was a time when all international telex was monitored by humans, but that of course telex volume was very low in the 60s through the 80s compared to email today.
Telex?
The NSA’s first mass spying program! They were so cute back then.