Corpus design for Malay corpus-based speech synthesis system

Abstract

Problem statement: Speech corpus is one of the major components in corpus-based synthesis. The quality and coverage in speech corpus will affect the quality of synthesis speech sound. Approach: This study proposes a corpus design for Malay corpus-based speech synthesis system. This includes the study of design criteria in corpus-based speech synthesis, Malay corpus based database design and the concatenation engine in Malay corpus-based synthesis system. A set of 10 millions digital text corpuses for Malay language has been collected from Malay internet news. This text corpus had been analyzed using word frequency count to find out all high frequency words to be used for designing the sentences for speech corpus. Results: Altogether 381 sentences for speech corpus had been designed using 70% of high frequency words from 10 million text corpus. It consists of 16826 phoneme units and the total storage size is 37.6Mb. All the phone units are phonetically transcribed to preserve the phonetic context of its origin that will be used for phonetic context unit. This speech corpus had been labeled at phoneme level and used for variable length continuous phoneme based concatenation. Speech corpus is one of the major components in corpus-based synthesis. The quality and coverage in speech corpus will affect the quality of synthesized speech sound. Conclusion/Recommendation: This study has proposed a platform for designing speech corpus especially for Malay Text to Speech which can be further enhanced to support more coverage and higher naturalness of synthetic speech.