Selection of Text Script for Text-to-speech Synthesis

Y. Feng (PRC)


speech synthesis, text corpus, greedy algorithm, KullbackLeibler measure


A large text corpus (more than 3 million sentences) is collected, transcribed phonetically and represented by context-varied syllable vectors (CSV). In order to cover most frequent units and basic units of CSV, this set covering problem is solved efficiently using a modified greedy algorithm taking unit frequency and sentence length into account. Compared with another method based on Kullback-Leibler measure, greedy algorithm shows better coverage and efficiency. The generated text scripts are about 13 thousand sentences. The set of selected sentences covers all desired units and 33.9 instances per unit are condensed to 1.5 instances per unit.

Important Links:

Go Back