AUTOMATIC CONVERSION OF TEXTUAL INFORMATION INTO SPEECH

Milan Sečujski

Vlado Delić

Apstrakt:

The task of automatic conversion of text into speech (Text-to-Speech Synthesis – TTS) is to convert input information in textual form into a speech signal understandable by humans. Synthesized speech, thus, has to sound natural, i.e., to possess the intonation characteristic of actual human speech. Speech synthesizers are used within various applications in communication networks, but also have a number of other applications of interest to the military industry, as they represent an indispensable component of human-machine spoken dialogue.

The paper describes the speech synthesizer for Serbian developed at the Faculty of Technical Sciences, Novi Sad. This system yields highly natural synthetic speech owing to accurate morphologic annotation and sentence accentuation, as well as state-of-the-art statistic models used for prediction of sentence intonation and other elements not explicitly given in the text but nevertheless critically important for the quality of synthetic speech. Speech signal synthesis is carried out by concatenation of speech segments from a 12-hour speech database (out of which only 4 hours are used in the current version of the system), where the string of segments to be used for synthesis is selected based on phonetic and prosodic specifications of the utterance to be synthesized. The solution developed is of the highest quality in the area where South Slavic languages are spoken, and the research into the implementation of this system into spoken dialogue systems in Serbian is currently under way.

Key words: speech synthesis, speech synthesizer, speech signal, text processing, communication interphase, man-machine dialogue, Serbian language.