SVOX: The Implementation of a Text-to-Speech System for German

This PhD thesis describes the implementation of the SVOX text-to-speech system for German. Several researchers participated in the design and realization of this system, whose contributions are summarized in the present thesis. The author's own developments in terms of system design and processing methods are described in detail: (i) The conversion of abbreviations, numbers, and novel words into pronunciation form, which is usually done in an independent text preprocessing and a letter-to-sound module in other systems, was fully integrated into the lexicon-based morphological and syntactic analysis of the SVOX system in a straightforward manner. The morpho-syntactic analysis and all grapheme-to-phoneme mapping is carried out in a framework of DCGs (definite clause grammars) and two-level finite-state morphology rules. (ii) Accentuation and prosodic phrasing were designed according to algorithms given in the linguistic literature. (iii) Fundamental frequency is generated by means of recurrent neural networks which were trained to map phonological data (phoneme sequence, accents, and phrase boundaries) onto corresponding natural fundamental frequency contours.

Due to the application of very general processing methods, the SVOX system is comparatively simple and flexible. In order to encourage the realization of a linguistically plausible architecture, real-time processing was not made a design restriction. However, the current SVOX system is able to synthesize simple texts in real time on newer workstations (VAXstation, SPARCstation).