In developing a text-to-speech system, it is well known that the accuracy of information extracted from a text is crucial to produce high quality synthesized speech. In this paper, a Persian text to speech system is studied. The system uses speech waveform concatenation method that comparatively mature in text-to-speech synthesis This paper describes the innovation introduced into the text analyzer module in a text-to-speech
system. In this analyzer, a probabilistic model is used along with a database for text to phoneme conversion. We call this probabilistic model Smooth Ergodic Hidden Markov Model (SEHMM and show that is an effective choice for text to speech applications