You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
49 lines
2.7 KiB
49 lines
2.7 KiB
The phonological component
|
|
|
|
questions. If clauses are conjoined, a syntactic symbol is placed just before the
|
|
conjunction. If a comma could be placed in the orthographic rendition of the
|
|
desired utterance, then the syntactic comma symbol “,” should be inserted. Syn-
|
|
tactic commas are treated as full clause boundaries in the rules; they are used to list
|
|
a series of items and to otherwise break up larger units into chunks in order to
|
|
facilitate perceptual processing.
|
|
|
|
The end of a noun phrase is indicated by )N. Segments in the syllable prior
|
|
to a syntactic boundary are lengthened. Based on the results of Carlson et al.
|
|
(1979), an exception is suggested in that any )~ following a noun phrase that con-
|
|
tains only one primary-stressed content word should be erased. The NP + VP is
|
|
then spoken as a single phonological phrase with no internal phrase-final lengthen-
|
|
ing and no fall-rise FO contour to set off the noun phrase from the verb phrase.
|
|
|
|
8.3 Comparison between ideal synthesis input and system performance
|
|
An example of the output of the analysis routines of MITalk is presented in Sec-
|
|
tion 8.7 at the end of this chapter. Examples where the analysis routines made an
|
|
“error” are underlined in Section 8.7, and the seriousness of the error is indicated
|
|
by a footnote for those errors deemed detrimental to perception. The word “error”
|
|
is put in quotation marks to emphasize that an error made by an analysis routine
|
|
need not be an error in some abstract linguistic sense, but only an error in the sense
|
|
that the symbol is not the one that is desired by the synthesis routines.
|
|
|
|
There are over 200 words in the sample text of Section 8.7 and over 1000
|
|
phonetic segments.
|
|
|
|
8.3.1 Phonetic transcription “errors”
|
|
|
|
There are 25 phonetic transcription errors, all minor, most of which concern the
|
|
difference between “I” and schwa. There do not seem to be serious problems with
|
|
the letter-to-sound rules, in part because they are rarely activated, i.e., about five
|
|
percent of the time. The rate at which phonetic errors are produced during MITalk
|
|
analysis, about one percent (i.e. about one word in twenty is in error in running
|
|
text), is quite good in comparison with text-to-speech systems that rely more
|
|
heavily on letter-to-sound rules. Sentence intelligibility and comprehension scores
|
|
are very high given the current analysis abilities.
|
|
|
|
8.3.2 Stress “errors”
|
|
|
|
There are 12 errors involving lexical stress assignment. Certain common words
|
|
such as “might” and “each” should be marked with primary lexical stress in the
|
|
lexicon because they almost always attract a certain amount of semantic focus, but
|
|
they are not currently assigned stress. Other words, such as “prerecorded”, are not
|
|
handled correctly by the morphological stress reassignment rules.
|
|
|
|
85
|