Vietnam Text To Speech System Front-end Text Analysis

Vietnamese is the official language of Vietnam. It is part of the Austro-Asiatic Mon Khmer group. Vietnamese is similar to Mandarin and each syllable can be marked with initial, final and tone. In order to develop the application system of Vietnamese Text-To-Speech (TTS), this thesis studies the Vietnamese text analysis method and explores ways of improving naturalness. The entire TTS system consists of a text analysis module (front-end processing) and a speech synthesis module (back-end processing). Text analysis module plays a vital role to improve the naturalness and intelligibility of output speech. Text analysis module involves text normalization, word segmentation, part-of-speech (POS) tagging and prosody phrase (L3) break prediction and so on.The main works in this thesis are as follows:1. Text normalization is realized by applying the method of rules and keywords. The experimental results show accuracy rate of97.2%.2. Automatic word segmentation is realized by applying the method of combination of the forward maximum matching method (FMM) and backward maximum matching method (BMM) and word frequency statistics. The experimental results show accuracy rate of97.24%, recall rate of91.35%and an F-1of94.20%3. POS tagging is realized by N-gram language modeling, and decoding process is realized by applying the Viterbi algorithm. The experimental result show accuracy rate is82.42%.4. Automatically predict Vietnamese L3break by using C4.5decision tree algorithm. The experimental results show accuracy rate of60.73%in close set, and the accuracy rate of open set is57.96%.