New prosodic phrasing model for Sinhala language

Abstract

This paper describes a new model of predicting prosodic phrase breaks in Sinhala language in order to improve the quality of the existing TTS Sinhala voices. In a Text To Speech (TTS) system, quality of the synthetic voice is mainly dependent on, how well its prosodic model is implemented. The prosodic model adjusts the phrasing and the pitch of the voice while applying suitable durations and tones for words and diphones. Out of these, phrasing and pitch of the voice carries much importance since appropriate phrase breaking helps to clearly understand the synthesis voice. In a real world scenario, when we speak a sentence, we automatically divide it to small segments and apply pauses at those breaks. Also the pitch of the voice gets lowered near a break and gets increased in the other segments automatically. But in a TTS system, we do not have that advantage and therefore need to be precised with the phrase breaks. Otherwise it will create wrong meanings as well as producing unnatural speech. Existing Sinhala TTS systems lacks proper prosody implementations and hence difficult to understand when it reads, especially long sentences. This issue can be overcome by applying a suitable phrase breaking technique.