Development of an hmm based bilingual synthesizer for tamil and indian english by merging acoustically similar phonemes

Abstract

A text-to-speech (TTS) synthesis system aims to generate intelligible and natural speech corresponding to any given text. Such a system could be used to synthesize speech in a single language or in countries like India, where people tend to interleave English words with their native language in written and spoken forms (mixed-language sentences), it would be desirable to develop a synthesizer that is capable of synthesizing monolingual and mixed-language sentences. In this regard, the current work focuses on developing a bilingual synthesizer for the Indian language, Tamil and Indian English. Initially, a phone-mapping approach is used to develop a mixed-language synthesizer, trained only on Tamil data, by mapping English phonemes to the perceptually similar Tamil phonemes. It is observed that this synthesizer replaces all English phonemes by a corresponding Tamil phoneme, and some phonemes, specifically those that are unique to English are mispronounced and unintelligible. Therefore, it would be preferable to develop a bilingual synthesizer with data from both languages. In such a case, the two straight-forward approaches to develop a bilingual synthesizer are (a) building separate synthesizer for each language and combining them or (b) merging the perceptually similar phonemes pertaining to the languages. In the former approach, switching from one language to the other, primarily in the form of a change in energy and accent at the language switching points is perceived, which could be annoying when mixed-language sentences are synthesized. With the latter approach, while there is no perceivable language switching, an influence of English on Tamil phonemes is perceived, which is undesirable and affects the intelligibility of the synthesized speech. newline newline newline