Sinhala Text to speech with unit selection

Abstract

This thesis describes an attempt to generate a more natural-sounding Sinhala Text-to- Speech system using the unit selection mechanism. The existing Sinhala Text-To-Speech engine, developed by UCSC LTRL (University of Colombo School of Computing, Language Technology Research Lab) has some drawbacks regarding the naturalness of the voice and correct pronunciation. This lack of naturalness can be addressed, with a limited set of units from which speech is chosen (from whole phrase down to diphone or even smaller). Sinhala unit selection has not been attempted yet for a text to speech application. Therefore, successful attainment of this research would be immensely helpful to the whole of the Sinhala speaking community and especially to those who wish to read in an eyes-free environment, the visually impaired, those who have learning disabilities and even those who are learning Sinhala as a second language. A Unit-Selection speech database requisites a high volume of recorded speech, at least 100 hours, for maximum naturalness of voice to be attained [27]. Also, speech recording needs to be done in a place void of interference in order to achieve the best quality possible. As most of the existing research had utilized the Festival framework to develop Sinhala Text-to-Speech, MARY, which is an open-source, multilingual Text-to- Speech Synthesis platform written in Java, was used to implement this research. A study on the MARY framework was carried out to understand how it works. Creation of the unit-selection Sinhala text to speech system on MARY was achieved in two steps: building of language support for Sinhala and building of Sinhala unit-selection voice. In order to accomplish this task, sample text files from the Sinhala News Corpus from UCSC?s LTRL were used to create the text database, and fresh voice recordings were made with a non-professional, native, female speaker. Results proved that unit selection successfully improved the naturalness of Sinhala synthesized speech.