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.
50 lines
2.6 KiB
50 lines
2.6 KiB
From text to speech: The MITalk system
|
|
|
|
8.3.3 Morpheme boundary problems
|
|
|
|
The morpheme boundary symbol * is used in the synthesis rules to prevent words
|
|
like back*ache from having a strongly aspirated medial kx. However, in a word
|
|
such as applic*ation, a restructuring of syllable boundaries is desirable so that the
|
|
medial Kx is strongly aspirated. In the present rule system it is not, since the * is
|
|
in the way. Perhaps the morpheme boundary symbol should be deleted between a
|
|
root and bound suffix (but not between two root morphemes). In other related
|
|
cases, the boundary prevents desired resyllabification processes so that
|
|
automatic*al*ly comes out as a six-syllable word, rather than the more normal 20
|
|
DX - AX - MM 'AE DX - IH - KK LL IY.
|
|
|
|
8.3.4 Syntactic “errors”
|
|
|
|
There are a rather large number of syntactic "errors” involving the incorrect assign-
|
|
ment of phrase and clause boundary locations. There are seven examples of a
|
|
missing end-of-phrase )N symbol, one missing end-of-clause )c symbol, and 17
|
|
cases where an end-of-clause symbol was incorrectly inserted between words with
|
|
the intent to break up longer phrasal units. This had undesirable perceptual im-
|
|
plications. The current algorithm intentionally adds extra clause boundary sym-
|
|
bols in order to break up the synthesis into smaller groups of words set off by
|
|
pauses and intonation breaks. These extra pauses were added because the com-
|
|
puter seemed to be able to go for long stretches without “pausing for breath”. The
|
|
trade-off between adding breath pauses to break the speech up into fewer process-
|
|
ing chunks versus insertion of a break at a syntactically unacceptable place has yet
|
|
to be optimized.
|
|
|
|
8.3.5 Summary
|
|
|
|
Of the analysis errors that were encountered in this admittedly difficult passage,
|
|
most of the phonetic, stress, and phonological rule errors are easily correctable.
|
|
However, only a few of the syntax errors can be fixed by straightforward debug-
|
|
ging techniques. The most serious limitation of text-to-speech analysis routines
|
|
seems to be in the area of automatic syntactic analysis. Still, the intelligibility and
|
|
comprehension results to be presented in Chapter 13 indicate very encouraging
|
|
overall system performance.
|
|
|
|
8.4 Stress rules
|
|
|
|
The phonological component assigns a feature Stress (value = 0 or 1) to each
|
|
phonetic segment in the output string. The default value is O (unstressed). Vowels
|
|
preceded by a stress symbol (°, ", or !) in the input are assigned a value of 1. Con-
|
|
sonants preceding a stressed vowel are also assigned a value of 1 if they are in the
|
|
|
|
same morpheme and if they form an acceptable word-initial consonant cluster.
|
|
|
|
86
|