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.
39 lines
2.4 KiB
39 lines
2.4 KiB
3
|
|
|
|
Morphological analysis
|
|
|
|
3.1 Overview
|
|
|
|
MITalk is designed to convert unrestricted English text into a synthetic speech
|
|
waveform. In the initial analysis phase, text character strings are converted to a
|
|
narrow phonetic transcription consisting of phonetic symbols and prosodic
|
|
markers. While the output unit types are thus specified, the question remains as to
|
|
the type of unit to be used with the input character string. Since there is an infinite
|
|
number of possible English sentences, it is not possible to store all English sen-
|
|
tences and their corresponding phonetic transcriptions in a form suitable for the
|
|
synthesis phase of MITalk. The next smaller unit recognizable from the input
|
|
string is the word. The number of English words is large, but bounded, so one
|
|
might consider use of a word lexicon which would contain the spelling and
|
|
phonetic transcription (together with part-of-speech information) for all English
|
|
words. Aside from the size of this dictionary, there are several attractive features
|
|
of this approach. Some form of dictionary must be used to provide pronunciations
|
|
for exceptions to other mechanisms (e.g. rules) used to derive pronunciations.
|
|
These arise in part from foreign words that have retained the pronunciation of their
|
|
language of origin (e.g. parfait and tortilla). Furthermore, all mechanisms
|
|
derived thus far for the conversion of letter strings to phonetic segment labels
|
|
provide some errors, and it seems to be inherent in natural languages that no for-
|
|
mal means derived for the representation of their structure has covered all ob-
|
|
served forms without error. An interesting class of exceptional pronunciation
|
|
arises for high-frequency words. Initial th is pronounced as a voiceless fricative in
|
|
many words (thin, thesis, thimble) but for very frequent words, such as the short
|
|
function words (the, this, there, these, those, etc.), it is pronounced in a voiced
|
|
manner. Similarly, f is always pronounced as an unvoiced fricative, except for the
|
|
single case of. In words such as shave and behave, the final silent e has the effect
|
|
of lengthening or tensing the preceding vowel, but in the frequent word have this
|
|
is not the case. Finally, the final s in atlas and canvas is unvoiced, but for the
|
|
function words (is, was, has) it is voiced. It thus appears that these high-frequency
|
|
words should be placed in an exceptions dictionary if a set of rules is to be used for
|
|
converting letter strings to phonetic segment labels.
|
|
|
|
23
|