11-20-96

	Minimal progress has been made with our project:
After researching the issues involved in speech recognition, we
decided to begin with vowel recognition.  This is because many
components of speech contain sharp, discountinious components
(like the s in components).  Therefore, we will resolve vowel
recognition and continue from there.
	For vowel recognition we recorded a 1 second sample of
each member in our group saying a, e, o and u.  We "normalized"
the samples to remove any DC components and also choose the sample
vector lenght in matlab to be a power of 2 (allowing the wavelet
transform to be fully complete an integer number of levels).  Our
plan was to take the vocal samples, perform a wavelet transform
and compare any test vocal sounds to the "training" data to
determine who the speaker is.  When we proceeded to do these
tests, the data produced by the wavelet transform didn't
correspond to what we had predicted.  Feeding sample signals
like an impulse and chirp into the "dwt" (discrete wavelet
transform) command produced obviously WRONG results.  Likewise, 
our audio signals which clearly showed low frequency components 
in the fft spectrum showed HIGH frequency components in the 
Wavelet domain.
	After a lenghty discussion with Jan Odegard on Thursday
10-21, we determined that the matlab code was malfunctioning for 
owlnet users.  On Monday 11-25, I aquired an account on the ece
machine Jazz.  I immediately tested the matlab toolbox and it
appears to work correctly.  Our gang plans on conducting all future 
experiments on Jazz.