Synthesis Unit Selection and Word Prosody Adjustment in a Chinese Text-to-Speech System

Abstract

This paper proposes a Chinese text-to-speech conversion system which focuses on the selection of synthesis unit and word prosodic pattern. In order to take into account both prosodic and phonetic information between concatenated syllables, synthesis units are selected from a large continuous speech database. On the other hand, since word is the basic rhythmical pronunciation unit, a word-based prosody modification approach is proposed. From observation, it appears that the prosodic properties of a Chinese word is generally affected by the tone combination, word length, part of speech of the word, and word position in a sentence. Consequently, a word prosody database is established based on above linguistic features. Each word pattern in the word prosody database contains linguistic features and prosodic patterns of the word. Moreover, the cubic spline curve fitting method is adopted to describe the prosodic pattern. A word prosody selection method was proposed to select appropriate word prosodic pattern from the word prosody database. According to the result of listening test, and adequate sequence of synthesis units and word prosodic patterns was obtained.