A syllable based Chinese synthesis system

Viac o knihe

The automatic conversion of written text into spoken language, known as text-to-speech (TTS) synthesis, is increasingly vital in human-computer interaction. At the Dresden University of Technology, a multilingual TTS system called DreSS was developed, and this thesis focuses on the Mandarin Chinese component of DreSS. The process involves excising syllables from spoken utterances and connecting them to create the desired speech signal, addressing two key research areas: inventory structure and prosodic parameter control. To enhance naturalness in synthesis, two innovative approaches were implemented. First, a coarticulation-oriented inventory was constructed to model coarticulatory effects across syllable boundaries. Second, a rule-based sentence intonation was superimposed on the inherent tone structures to preserve the natural tone contours of Chinese syllables. This strategy minimized concatenation points and reduced tonal contour modifications, resulting in a more natural Mandarin Chinese synthesis. Furthermore, the thesis explores two essential components of prosody control in Mandarin speech synthesis: duration modeling and tonal concatenation, utilizing a natural speech database for analysis. The findings, along with suitable models for duration and tonal coarticulation, significantly enhance the synthesis system's performance.