from-text-to-speech-the-mitalk-system/pages-txt/111.txt

The prosodic component

ness and intelligibility ratings of sentences synthesized by these rules are very
similar to ratings of the same sentences synthesized using durations obtained from
a natural recording.

Complete durational rule systems exist for English (Coker et al., 1973) and
Swedish (Carlson and Granstrom, 1976). (We have borrowed heavily from the
elegant rule system of Lindblom and Rapp that was augmented and implemented
by Carlson and Granstrom.) Partial rule systems have also been proposed for
vowels (Umeda, 1975; Liberman, 1977) and for consonants (Umeda, 1977). The
rules contained in these systems are similar (not surprisingly), but there are many
ways to generalize from the available data. For example, Coker et al. (1973) rely
heavily on multiple stress levels conditioned by syntactic category (verbs have less
stress than nouns) and conditioned by word frequency (common words and words
that are repeated in a discourse are reduced in stress). Liberman (1977) includes
rules related to rhythm and isochronous principles. Neither of these kinds of rules
are incorporated explicitly in our system, but we do achieve partial isochrony
through rules that shorten unstressed syllables and consonant clusters (see Carlson
et al., 1979). For quantification, we capture durational differences between nouns
and verbs by phrase-final lengthening, and we permit the use of the emphasis sym-
bol “!* in the input to capture word frequency and discourse expectancy effects in
a binary fashion. |

Therefore, it may never be possible to make absolute judgements concerning
which rule system is theoretically correct. Effort should rather be directed at sys-
tematic optimization of a particular rule system, e.g., one that starts with a linguis-
tically motivated framework for how to represent an input sentence and draws on
both speech production data and perceptual constraints to formulate a simple set of
rules as a starting point.

99