|
|
Preface
|
|
|
|
|
|
necessary data to support the extensions and refinements to the morph analysis
|
|
|
routines. Subsequent to the initial construction of the lexicon, an elaborate editing
|
|
|
of all entries was made by M. S. Hunnicutt. This led to substantial improvements
|
|
|
in the system’s overall performance.
|
|
|
|
|
|
When words could not be found in the morph lexicon, or could not be
|
|
|
analyzed into morphs from the lexicon, letter-to-sound rules were utilized. Prior to
|
|
|
the MITalk research, letter-to-sound rules had been proposed to cover the entire
|
|
|
language. But, with MITalk, it was realized that high-frequency function words
|
|
|
often violate perspicuous forms of these rules, and that such letter-to-sound rules
|
|
|
do not span morph boundaries. Based on these observations, a complementary set
|
|
|
of letter-to-sound rules could be introduced into MITalk, but these rules would not
|
|
|
be used unless morph analysis failed. Realizing this fact, affix stripping was util-
|
|
|
ized, and the more reliable consonants were converted first, leaving the vowels for
|
|
|
last. This approach was proposed by J. Allen, and extensive sets of these rules
|
|
|
were developed by M. S. Hunnicutt working with F. X. Carroll. Several sets of
|
|
|
these rules were developed and elaborately tested. In addition, in the late *60s at
|
|
|
MIT, there was great interest in lexical stress and phonological rules for this pur-
|
|
|
pose which were initially developed by M. Halle and N. Chomsky. These rules
|
|
|
were reformulated and extended to include the effect of affixes. This was the first
|
|
|
time that lexical stress rules had been used in a text-to-speech system. The
|
|
|
development of rules for this purpose, along with their unification with the letter-
|
|
|
to-sound rules, was accomplished by M. S. Hunnicutt. In addition, the text
|
|
|
preprocessing rules were also provided by M. S. Hunnicutt, as well as the routines
|
|
|
for morphophonemics and stress adjustment used in conjunction with the morph
|
|
|
analysis.
|
|
|
|
|
|
In a 1968 doctoral thesis, J. Allen developed a parsing methodology for use in
|
|
|
a text-to-speech system, with particular emphasis on the computation of necessary
|
|
|
syntactic markers to specify prosodic comrelates. This parsing strategy led to the
|
|
|
development of a phrase-level parser which avoided the complications of clause-
|
|
|
level parsing and the problems of syntactic ambiguity at that level, but also led to
|
|
|
the introduction of inaccuracy due to incomplete clause-level analyses. This ap-
|
|
|
proach was augmented and extended by P. L. Miller and C. J. Drake, and was
|
|
|
tested extensively in the context of the morph lexicon and analysis routines.
|
|
|
|
|
|
In light of the phonetic segment labels, stress marks, and syntactic markers
|
|
|
obtained by the previously mentioned programs, it was necessary to develop a
|
|
|
prosodic framework for the following phonemic synthesis. A durational
|
|
|
framework was developed by D. H. Klatt together with R. Carlson and
|
