Design and evaluation of a phonological phrase parser for Spanish text-to-speech

Abstract

This paper presents and evaluates a phonological phrase parser for a Spanish text-to-speech system. The parser consists of three stages: 1) lexical lookup, using a small dictionary (428 words); 2) preliminary phrase boundary placement, using a modification of Liberman and Church's (1992) function group parser; and 3) readjustment of phrase boundaries, using syllable count and punctuation. A corpus of 382 hand-parsed sentences (1,691 phrases) was used to evaluate the parser. The parser generated almost the same number of phrases (1,692) as the hand-parsed sentences with 70% (1,186) agreement. Suggestions for improving the parser's performance include the expansion of the verb lexicon, performing simple morphological analysis for verbs, and relaxing the syllable count in phrases before verb forms.

Assigning phrase breaks using CARTs for Basque TTS

“…In expression (1), overall score achieved by the tree is compared with the probability of having a non-break label in the data, eliminating the dependency on the structure of the data. If the algorithm does not insert any break, the value of the kappa statistic will be 0.…”

“…There are systems that place breaks by rule, taking into account a function/content word classification [1], and other systems use diverse methods of statistical analysis to insert breaks [2] [3]. As Basque is an agglutinative language very few function word exist, so traditional rules are not applicable in our case.…”