In recent years, speech synthesis has become an dynamic research area in the field of speech processing due to the usage of automated systems for spoken language interface. This paper address an innovative Tamil Text-to-Speech(TTS)synthesis system utilizing Hidden Semi- Markov Model (HSMM) to analyze sentiments from the speech output. Four different HSMM methods have been proposed to implement the above task. The sentiment-dependent (SD) modeling is the first method, utilize individual models trained for each emotion individually. Sentiment adaptation (SA) modeling is the second method, initially a model is trained using neutral speech, and the adaptation process are implemented to each emotion of the database. The third method is called sentiment-independent (SI) technique, is at first trained utilizing data from all the sentiment of the speech database which is based on an average emotion model. Subsequently, an adaptive model has been constructed for each emotions. The fourth method is called sentiment adaptive training(SAT), the average emotion model is trained with simultaneously normalization of the output and state duration distributions. These training methods are evaluated using a Tamil speech database which consists of four categories of speech, anger, joy, sadness and Disgust was used. To assess and compare the potential of the four approaches in synthesized sentimental speech an emotion recognition rate subjective test was performed. Among the four evaluated methods the sentiment adaptive training method gives the enhanced emotion recognition rates for all emotions.

HMM HSMM TTS