Diphone database development for a Romanian language TTS system

The development of the acoustic unit database represents one of the most difficult problems for concatenative speech synthesis systems. The quality of the database strongly influences the accuracy of the overall system and is a fundamental requirement for a high-natural synthetic speech. This paper describes the design and generation of the acoustic database for the Romanian language text-to-speech system that has been developed during the last years in the Speech Processing Laboratory of the Faculty of Electronics in Bucharest. Basic considerations that lead to allophone selection and also database implementation details are discussed. The paper emphasises the strategy used to improve the diphone speech synthesis by using an augmented unit inventory and choosing segment sequences that minimise the local spectral distortion at diphone boundaries. Evaluation experiments that confirmed the validity of our approach are finally presented