Datasheet

3 - 6
This syntactic and lexical analysis makes extensive use of general rules and
dictionaries containing exceptions. It takes contextual linguistic knowledge to detect
that "record" is a verb and "record" a noun. Dictionaries are again used to solve
such cases: suffixes such as "-ness", "-logy", "-ly" can help to determine the word
status.
From Isolated Sounds to Real Speech
The linguistic analysis and conversion of written symbols into sound equiva-
lents is by no means sufficient: these segments have to be concatenated success-
fully to form fluent, coherent speech. Who pronounces isolated sounds one after
the other with equal stress and volume for each one of them? You wouldn’t even
understand what was being said!
The phonetic module calculates precise speech patterns from the phoneme
output of the linguistic module. To assure the smooth, non “robotic” character of
the produced speech, the phonemes are replaced by sound segments originating
with native speakers. Although you are dealing with a “synthesized” voice, the
characteristics of the speaker’s original voice are still recognizable in the spoken
text.
This does not mean that individual phonemes are replaced in each case; sound
clusters are also used for a more natural coarticulation, for better transitions
between the individual phonemes. “Diphones” consisting of two sounds as /ai/ in
the word "hide" and “triphones” consisting of three sounds as /aie/ in "buyer" are
extensively used.
Other elements which must be added with native speaker-based models are a
natural prosody - intonation and insertion of pauses - of each phrase and the
length and pitch of each sound. (And that’s why we recommend you to enable
the multiline reading mode!) Here again, native speaker-based models are used
for optimal results.
In most languages, at least one important word in a sentence gets an accent.
Think of sentences such as "John went home" - where we mean it wasn’t Peter
- and "John went home" - and not elsewhere!
4chapter3.p65 2/27/2001, 8:47 AM6