A Concatenative Synthesis Based Speech Synthesiser for Hindi

Abstract

The document presents a speech synthesis system based upon hidden Markov models and decision trees. The development follows the approach generally implemented for speech recognition systems. A front end analysis of the database is done and the derived feature vectors are subject to phonetic hidden Markov models, which are then clustered based upon a decision-tree approach that models the context of the phones being considered. The individual phone segments form an acoustic leaf sequence which divides various contexts into equivalence classes. During synthesis, the phone sequence to be synthesised is translated into an acoustic leaf sequence by posing questions associated with different nodes of the decision tree to the immediate context of the phone. The sequence of terminal nodes thus obtained is amalgamated to obtain the desired pronunciation.