Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

Automation of the Training Procedures for Neural Networks Performing Multi-Lingual Grapheme to Phoneme Conversion

Horst-Udo Hain

SIEMENS AG, Corporate Technology, Munich, Germany

Any TTS system requires accurate grapheme tophoneme conversion routines due to the limited pro-nunciation dictionaries. Data driven approaches of-fer the possibility to train new languages withoutthe knowledge of rules. A big problem is to preparehuge databases for the training. In this paper theautomation of the data preparation and the patterngeneration for the training of a neural network is ad-dressed. This automation is language independent,and no native expert is required. The data prepara-tion, the generation of the training patterns and thetraining itself are done completely automatically.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Hain, Horst-Udo (1999): "Automation of the training procedures for neural networks performing multi-lingual grapheme to phoneme conversion", In EUROSPEECH'99, 2087-2090.