Abstract

This paper proposes a method that synthesizes any given word, by concatenating syllabic components held in a word database. This method is especially efficient in an automated interactive system that handles a huge number of words as variables used to fill the slots of recorded templates to provide customers with appropriate voice guidance. In our method, syllabic components are collected from each recorded word using only the mora position within the word and the word's length. Words with different mora length yield different syllabic components even if they have the same syllables in the same mora positions. To synthesize a given word, components that have the same position and the same mora length are selected from a syllabic database and are simply concatenated to create the voice output required. The proposed method assumes that just the mora position and the mora length represent enough of the prosodic features to realize synthesis at the word level. Neither signal processing nor consideration of the prosodic factors such as fundamental frequency, power, duration and others are needed in either the collecting or synthesis stage. An experiment study to synthesize Japanese city and town names with five moras by the proposed method was carried out. Competitive opinion scores were obtained by the synthesized voices compared with those obtained by actual voices. By our estimation, only 17,000 city or town names need to be recorded to cover the remaining names with 3, 4, and 5 mora, which totals about 105,000.