HMM-based speech synthesis system (HTS) for Gujarati language.

Abstract

Hidden Markov Models (HMM) have been applied successfully to Automatic Speech Recognition (ASR) problems and are currently applied in speech synthesis applications. In this thesis, HMM-based Speech Synthesis System (HTS) for TTS is understood in detail and applied to Gujarati language. In particular, for HTS implementation, issues related to characteristics of Gujarati language are identified and their solutions are also discussed in this work. Furthermore, classification of Gujarati characters have also been done for preparation of question set. We have done comparison of conventional Baum-Welch algorithm which is the ExpectationMaximization (EM) algorithm for discrete finite state HMMs and DeterministicAnnealingExpectation Maximization (DAEM) algorithm in the context of HTS as applied to Gujarati language. Detail study on derivation of EM and DAEM algorithms alongwith examples have also been presented. Based on the subjective evaluations of HTS system, it was observed that HTS voices developed for Gujarati language has very high intelligibility. We found that as amount of training data increases, MOS of HTS voice improves. From results, we found that for 70.5 % of time people have preferred DAEM-based HTS than EM-based HTS. We also found that runtime memory of HTS developed for Gujarati language is in the order of few megabytes. Hence, it can be very useful for applications which suffer from memory limitations, such as mobiles and tablets.