“EYES FOR THE BLIND” - GENERATING HIGH QUALITY TEXT-TO-SPEECH VOICES FOR SINHALA

Abstract

This research represents a study carried out to compare different Text to Speech (TTS) modelling techniques used in developing Sinhala Text to Speech systems. In the area of speech research there are different modelling techniques. From that we can find Sinhala TTS systems based on formant synthesis, diphone concatenation and unit selection approach. HMM based modelling is another popular technique which have not attempted before for Sinhala language. Main objective of this study is to find out the best modelling technique for Sinhala language. Therefore formant synthesis, diphone concatenation, unit selection approach and HMM based modelling is evaluated in order to achieve the research question. As the first step of the study an HMM based TTS system is developed. Later these four modelling techniques are analyzed using user evaluations. Three quality attributes, as speech quality, naturalness and intelligibility are considered in the analysis with the user preference. As the results stands unit selection approach is the best technique in overall values of three quality attributes measured. Having a 75.20% of speech quality, a 78.20% of naturalness and an 86.20% of intelligibility. When it comes to the user preference 50% of the users preferred unit selection approach. With overall results diphone concatenation approach had the second best values.