CMU SPHINX TESTS (PRIOR TO MAY 1988)

TEST MATERIAL: The test set used by Lee in his dissertation
research with SPHINX consisted of a synthesis of the Speaker
Independent test material used in the March '87 and a portion of
the Speaker Independent test material designated for use in the 
October '87 tests.  It consisted, in all, of 150 sentence
utterances from 15 speakers.  

SPEAKER INDEPENDENT SYSTEM TRAINING: Lee defined a system training
set consisting of all 40 sentence utterances from each of the 80
speakers of the Speaker Independent Training portion of the
Resource Management Corpus, plus another 40 utterances (including
the "rapid adaptation" sentence ustterances) from 25 of the
speakers in the Speaker Independent Evaluation Test portion.   

LIVE TALKERS: No "live talker tests" were included in Lee's SPHINX
research. 

"HOMOPHONES CORRECT" OPTION IN SCORING: In Lee's dissertation,
"confusions between homonyms are considered correct recognition
when no language model is used, because homonyms have the itentical
dictionary entries, and are indistinguishable.  When a grammar is
used, homonym confusions are counted as [ordinary] sutstitution
errors."
 
STANDARD WORD-PAIR GRAMMAR: Lee made use of two grammars in
addition to the "no grammar" condition: (1) the non-probilistic
"word-pair grammar, which is a simple grammar that specifies only
the list of words that can legally follow any given word" (proposed
by BBN), and (2) the "bigram grammar", which is similar to the
word-pair grammar "except that the probabilities of tag transitions
are now estimated instead of assumed to be equal".