00100	As an aftermath of the very successful CMU  Segmentation  meeting  it
00200	has occurred to me that it would be very helpful if we could all have
00300	detailed  comparisons  of  the  way  different  components   of   our
00400	segmentation and labeling schemes behave. To this end I would like to
00500	propose that we first compare the results that we  get  in  labelling
00600	segments  as  to  whether they are voiced or not, or perhaps we might
00700	compare the results when we try to put each segment into one of  four
00800	classes, these being (1) Voiced but not with frication, (2) Frication
00900	without voicing, (3) voiced-fricative, and Stops, which would include
01000	pauses, unvoiced stops, and voiced stops.
01100	
01200	It  happens  that  we are able to make a very rapid evaluation of the
01300	usefulness of any subset of two  different  parameters  in  terms  of
01400	their  utility  in identifying any set of features as found in the 31
01500	test sentences. We can supply this evaluation via the net if  we  are
01600	supplied  with a file of the parameters with an identifying label for
01700	successive samples of the acoustic signal,  given  the  locations  of
01800	these samples, for the 31 test sentences. It simplifies our procedure
01900	if these parameters have values in the range of 0 to 63. Our  results
02000	can  be  returned via the net as a file which will produce plots such
02100	as those that can be seen on  my  file  VOICED.PLT[NET,ALS]  which  i
02200	suggest  that  you all copy and print. Incidentally, we are not proud
02300	of these results and we are not supplying them as  samples  to  shoot
02400	at, but simply as something to show to form of the output. We realize
02500	that many of you have a single parameter to effect separations of the
02600	type  indicated.  However  these  composite  parameters  are probably
02700	obtained by some sort of  mathematical  relationship  between  actual
02800	measured parameters and our test would show the ideal separation that
02900	could be obtained from any two of these measured parameters if one is
03000	limited to an accuracy specified by 6 bits. The two dimensional plots
03100	that  we  obtain  are  two-dimensional  projections  made  from   the
03200	n-dimensional  space  represented by all of the measurements that are
03300	made.