A novel Bengali text-to-speech system to achieve both intelligibility and naturalness using a small voice database

One of the key objectives for a Text-to-Speech (TTS) system is to achieve intelligibility and naturalness. However, designing such system using a small memory footprint becomes a challenging task. This paper proposes a concatenative Bengali TTS system using very small voice database. It mainly uses syllables as the building blocks of the concatenation processes. To limit database size, frequently-used context-sensitive syllables are identified and stored in a database. The similarity between target syllable and its closest matched syllable in the database is calculated. If it is higher than a certain threshold, matched syllable is used for concatenation; else the system switches to the diphone-based synthesis. This dual synthesizing model ensures that the system will never suffer from completeness issue. Since, maximum portion of the artificial speech comes from pre-recorded syllables, which contains higher amount of prosodic information, speech output sounds quite natural. The experimental results also prove its correctness.