5th International Conference on Spoken Language Processing
The Bell Labs text-to-speech synthesis system for French is part of a multilingual effort for text-to-speech generation. The text analysis component consists of four main parts: the morphological analysis module, the language models, the grapheme-to-phoneme conversion rules, and the prosodic module. The system is built in a pipeline architecture, the output of which feeds the subsequent synthesis modules. The originality of this work lies in the fact that we use weighted finite-state transducer technology to process the entire analysis of the French system. Moreover, the implementation not only accounts for most orthographic representations, such as numerals, abbreviations, dates, currencies, etc, but we also solve the hard questions of French liaison, mute e, and aspirated h using refined intermediate representations either in the form of traces or in the form of archigraphemes.
Bibliographic reference. Tzoukermann, Evelyne (1998): "Text analysis for the bell labs French text-to-speech system", In ICSLP-1998, paper 0075.