Slashdot Mirror
Improving Noise Analysis with the Sound of Silence
Roland Piquepaille writes "Researchers at Rockefeller University have built a mathematical method and written an algorithm based on the way our ears process sound that provides a better way to analyze noise than current methods. Not only is their algorithm faster and more accurate than previous ones used in speech recognition or in seismic analysis, it's also based on a very non-intuitive fact: they know what a sound was by knowing when there was no sound. 'In other words, their pictures were being determined not by where there was volume, but where there was silence.' The researchers think that their algorithm can be used in many applications and that it will soon give computers the same acuity as human ears. Read more for additional references and pictures about this algorithm."
Tag this as "rolandpiquepaillespam"
Heck...a little bit more and you've got a Simon and Garfunkel song. Perhaps you could even explain why the words of the prophets are written on the subway wall. But I digress.
The summary is completely useless, and the article isn't much better. After reading the description 3 times, I figured out the graph, at least. X-axis is time, Y-axis is frequency, and color is amplitude, so it's essentially a time dependent power spectrum density (PSD) with anything above a cutoff amplitude shown in black. I believe the Navy uses a variation of this called a waterfall to help interpret sonar sounds. I got stuck again, though, reading their description of the sample.
Aside from the apparent infinite-energy contradiction if this were true, the graph clearly shows that the signal is both frequency and time dependendent. Obviously, that's the case they ultimately have to deal with to apply this method, but the article suggests otherwise.
As for what they're actually doing, presumably, instead of operating on a set of data that includes time, frequency, and amplitude, they are cutting it down to time, frequency, and sound/no sound. This would cut the data size by the amplitude resolution (eg, 1/16th for a 16 bit amplitude sampling). This must assume that amplitude is irrelevant to the sound, which based on my (limited) experience working with PSD's, I'm skeptical of. Perhaps the odds of getting the same digital time/frequency data for two different sources is low enough that this can be ignored, much like the likelihood of two data sets yielding the same MD5 sum is non-zero but insignificant.