All-Prosodic Speech Synthesis
Arthur Dirksen
John Coleman
This document presents audio demos complementing our paper All-prosodic
speech synthesis, which appears in:
J.P.H. van Santen, R.W. Sproat, J.P. Olive, & J. Hirschberg
(eds.)
Progress in Text-to-Speech Synthesis, New York, Springer
Verlag. 91-108.
The audio demos are 16-bit .wav files, sampled at 11025 Hz, generated
by IPOX.
Demo 1 - Coarticulation
Spreading of vocalic place features in the phonology is reflected in phonetic
interpretation by subtle differences in frication spectra associated with
/s/:
Demo 2 - Syllable overlap
Three versions of /bot@l/ bottle, with ambisyllabic
/t/,
generated with different amounts of syllable overlap:
Demo 3 - Ambisyllabicity
Intervocalic clusters are parsed with maximal ambisyllabicity. In
the following words, the bracketed clusters are ambisyllabic:
Note that /t/ is aspirated in winter, but not in system.
Demo 4 - Syllable compression I
Again, three versions of /bot@l/ bottle, this time with different
amounts of compression for the first and second syllable:
Note that when the second syllable /t@l/ is compressed to 62%, the
vowel is almost fully eclipsed, creating the impression of a syllabic
sonorant.
Demo 5 - Syllable compression II
Two versions of /s^powz/ suppose, with different amounts
of compression for the unstressed prefix /s^p/:
-
60%
(reduced vowel)
-
52%
(vowel eclipsed)
Demo 6 - Syllable compression III
A segmental analysis of vowel elision would seem to predict that s'pport
is
phonetically identical to sport . Our analysis in terms of syllable
compression correctly predicts subtle (and less subtle) differences:
Note that /p/ is aspirated in s'pport, but not in sport.
Demo 7 - Syllable compression IV
The three words below have been synthesized using full vowels (/ow/
as in blow, /o/ as in pot, and /a/ as in sad
)
in analysis as well as phonetic interpretation. By varying syllable compression
in accordance with metrical-prosodic structure, we obtain the expected
alternations between full and reduced vowels.
Demo 8 - Connected speech
Our first attempt at generation of a full sentence with IPOX.
Arthur Dirksen, John Coleman
February 1995, revised June 2000