AN ANN-BASED METHOD TO IMPROVE THE PHONETIC TRANSCRIPTION AND PROSODY MODULES OF A TTS SYSTEM FOR THE ROMANIAN LANGUAGE

摘要

High quality Text-to-Speech applications require input text pre-processing in order to provide the speech synthesizer with reliable information regarding both phoneme parameters and prosody. Part of this mandatory pre-processing refers to phonetic transcription, syllable delimitation and stressed syllable identification in words. The translation of letters (or graphemes) strings into sequences of phonemes may be done almost entirely by rule-based methods, built on knowledge extracted from phonetic and phonologic rules, because the Romanian language has a phonetic character. Nevertheless, there may exist situations that cannot be entirely characterized by explicit rules, namely those of the two/three consecutive vowels pronounced either in a diphthong/triphthong, or in a hiatus. We describe our approach for Romanian phonetic transcription and stress assignation by use of two ANN sub-modules. The first sub-module implements the word hyphenation by an automatic learning method, solving the decision diphthong/triphthong versus hiatus (including semivowels identification), also covering the exceptions to existing syllabication rules. The output supplies information to the Letter-to-Sound module, for completing the phonetic transcription. The second ANN sub-module implements a solution for stressed syllable identification, as required in the prosody module.