Concerning Robert Walker's points

FROM: mclaren
TO: The new practical microtonality group
SUBJECT: Concerning Robert's points about the FFT

Robert Walker suggested using interpolation on short FFT
inputs. Other solutions include zero padding and resampling.
All these methods involve that sophisticated scientific method
known as "the wild-ass guess."
Whether you pad with zeroes or interpolate or perform some other mathematical prestidigitation, you are essentially crossing your fingers and hoping that niether the freuqency nor the magnitude of the spectral peak changed very much during the course of your analysis, or that if it did change it changed in a nice neat smooth way.
This is often not the case, as with speech, clarinet squawks, brass rips, violin tremolandi, etc., etc. etc.
The second and much more improtant problem with Robert Walker's suggestion involves the fact that as usual Robert Walker appears to be using the standard "toy input" frequency tones. These kinds of pure test tones prove delightful and lots of fun i nthe lab. We do not often encounter 'em in the real world.
For a better test and a more convincing proof of the general inadequancy of the Fourier Trasnform for analyzing real sounds in the real world, try the following:
Record 2 minutes of street ambience on a busy downtown streetcorner. Try to concoct a Fourier Trasnform based computer listening system which will reliably identify more than 75% of the sounds.
Second, try recording 1 minutes of typical Romantic symphony. Try to concoct a Fourier Trasnform basde computer listening stem which will reliably identify throughout the course of that 1-minute-long snippet the different orchestral instruments playing and the notes they play.
Eric Sheirer has already created a computer listening system which passes test number 1. To do so, he had to abandon spectral methods of analysis such as the Fourier Trasnform in favor of time-based autocorrelation methods of analysis.
James Moorer tried to tackle task number 2. So did Pizscalski. They both failed. No one has yet succeeded at task number 2, though humans effortlessly accomplish it.
Bottom line?
If you want to prove me wrong, there's a simple and easy way to do it -- create a Fourier Trasnform based computer listening system which passes either of the above two real-world tests.
My prediction?
Robert Walker will not accomlish this. He will not accomlpish this because the brightest and best students at the most elite tehcnical universities ont he planet have barely managed to accomplish task 1, and then only by abandoning the Fourier Transform.
If Robert Walker does succeed in accomlishing either of these two tasks, by the way, he will have qualified for a doctorate from MIT or Stanford.
I look forward to hearing of your award ceremony, Robert. In the meantime, actions speak louder than words. Unless and until you can created such a real-world computer listening system, permit me to suggest that it is futile to speak about how useful the Fourier Transform is in dealing with real musical sounds in the real world.
--------------
--mclaren