Text analysis for French text to speech synthesis: optimization of the Bell Labs system

We report results on optimization experiments for text-to-speech systems (TTS) within the context of the multilingual platform at Bell Laboratories-Lucent Technologies. In the process of transferring the technology to the Lucent development unit, the research prototypes were too expensive both in space and processing time. Therefore, measures have been taken to remedy this situation and to transform the prototypes into highly robust systems. These measures involve examining all the areas of the system from a linguistic, phonetic, dictionary, and implementation standpoint. The purpose of this paper is to demonstrate the multiple areas in which optimizations of a text-to-speech front end have been achieved, showing old and new performance, and presenting the new optimized system. The newly optimized system has been reduced from 48.8 MB to 11 MB and runs more than twice as fast. The experience acquired in these optimizations can be generalized to other TTS systems.