Towards a high quality Arabic speech synthesis system based on neural networks and residual excited vocal tract model

Abstract

Text-to-speech conversion has traditionally been performed either by concatenating short samples of speech or by using rule-based systems to convert a phonetic representation of speech into an acoustic representation, which is then converted into speech. This paper describes a text-to-speech synthesis system for modern standard Arabic based on artificial neural networks and residual excited LPC coder. The networks offer a storage-efficient means of synthesis without the need for explicit rule enumeration. These neural networks require large prosodically labeled continuous speech databases in their training stage. As such databases are not available for the Arabic language, we have developed one for this purpose. Thus, we discuss various stages undertaken for this development process. In addition to interpolation capabilities of neural networks, a linear interpolation of the coder parameters is performed to create smooth transitions at segment boundaries. A residual-excited all pole vocal tract model and a prosodic-information synthesizer based on neural networks are also described in this paper.