Andrew Lentvorski wrote:
Oh, come on. I presumed that everybody on this list can do estimation.
Depends on the kind of estimation. I doubt many people here are capable of estimating how flops will be needed to analyze even one phone call much less millions. I certainly can't.
Okay, human voice has most of its information content below 4KHz. Sampling that requires at least 8Khz. Call it 10K samples/second.
It is easier for a person to understand a call recorded at higher sampling rate. Wouldn't it also be easier for a computer? Or wouldn't a voice recognition system be more accurat with a higher bit rate?
We'll work with 1 bytes per sample even though 2 bytes is probably better.
I think most digital phone systems only send one byte per sample anyway.
Now, we have to analyze that data. An FFT can be done in O(n log n). So we need 10^9 * log 10^9 flops. Or, roughly 10^10 flops continuously producing frequency bins.
I read that the latest FPU in a 3G x86 machine can do 24 GFLOPS. So nearly half a machine would be required, right?
Now, we have to analyze the DFT's and convert them to something useful.
What is a DFT? Discrete fourrier transform?
Hidden Markov Models seem to be on the order of O(n^2), so we go from 10^10 to 10^20. You need 10^10 (10 billion) computers operating at 10^10 flops (10 GHz) to chew through all of the data.
I wish I understood how a hidden markov model can be used for pattern recognition. I know that it can, but I have no clue how it works.
A little outside of even Google's ability to handle in real time. And I haven't even mentioned power.
But the problem you describe may not really be the problem they need to solve. They don't need to do general case speech recognition. If someone buys the adword "hamburger" for my geographic area then all they have to do is search for when I say hamburger. It seems like most big companies I call now have basic keyword recognition systems built into their PBX's which presumably have relatively modest hardware. So it seems like the problem may not be nearly as large as you describe.
This is why I'm so annoyed about the FBI wiretapping program. They're not analyzing the data in advance except for a *very* small number of people (and they're probably doing that with people). All they're doing is recording it for use in fishing expeditions afterward.
Indeed, that is annoying. -- Tracy R Reed Read my blog at http://ultraviolet.org Key fingerprint = D4A8 4860 535C ABF8 BA97 25A6 F4F2 1829 9615 02AD Non-GPG signed mail gets read only if I can find it among the spam. -- [email protected] http://www.kernel-panic.org/cgi-bin/mailman/listinfo/kplug-list
