From aer@research.att.com Sun Oct  6 12:03:45 1991
From: aer@research.att.com
Date: Fri, 4 Oct 91 16:56:59 EDT
To: jeff
Subject: Thesis Proposal

Jeff,
I scanned your thesis proposal.  It is basically OK, but will need
some reworking and editing.  Here are some random specific comments:

p.2, par. 3:  "cluster templates within a single state"  should be
"cluster vectors"

p.3, The definition and description of HMM's needs reworking.  You
don't mention the c's as model parameters.  I don't think it's
fair to state that the initial state probabilities "are not
very important".  (Certainly for a left-to-right model, the initial
probability is usually 1 for starting in state 1.)  You should
indicate why left-to-right topologies are natural for modelling
speech utterances.

p. 4, top.  We don't use state transition probabilities that are
equally likely for allowed transitions.

p. 4, middle.  What "simplifications" are you referring to?

p.6, par. 4.  You need to define and describe what you mean by
"fixed segmentation".

p.6 bottom.  Define what is meant by "positive" and "negative"
examples.

Aaron



