back to Contents
Institute
of Phonetic Sciences,
University
of Amsterdam,
Proceedings
22 (1998), 115 - 124
Rik
Jansen
[1])
,
Arjan J. van Hessen
1)
and Louis C.W. Pols
Abstract
To
improve pronunciation of the Fluent Dutch Text-To-Speech Synthesiser, two
pre-processors were built that try to detect problematic cases in input texts
and solve these automatically if possible. One pre-processor examines the
pronounceability of surnames and company names by checking whether their
initial and final two-letter combinations can be handled by the
grapheme-to-phoneme rules of the Fluency TTS system, and correcting those
automatically when and if possible. Also, common disambiguous abbreviations are
properly expanded. The second pre-processor tries to realise pronounceable
forms for numbers that do not have a straightforward pronunciation. Structural
and contextual information is used in an attempt to determine to what category
a number belongs, and each number is expanded according to the pronunciation
conventions of its category. It can be said that these pre-processors are a
useful aid in offline pronounceability examination (for names) and improvement
of performance at run-time (for numbers), although ambiguity and redundancy in
the input text illustrate the need for semantic and syntactic parsing to
approach human text interpretation skills.
back to Contents