Development of an automatic generator for multi-lingual knowledge bases for quick phoneme recognizer*

As a primary objective of the project, the conventional training methods have been developed into a novel method for automatic generation of knowledge bases for new languages based on already learned languages.

This procedure creates by splitting an existing knowledge base, the redistribution of acoustic models of an existing language and subsequent automatic adaptation of the models based on the heard speech patterns, an acoustic model of the target language during use.

Furthermore new features for version control and automated documentation of the processing steps have been created to avoid accidental editing errors and to permit editing by unqualified database users.

As part of the promotion project a new multilingual recognizer concept has been created and put into practice. The knowledge was brought from the field of speech synthesis and a novel discriminative training method has been developed which allows parallel use and the adaption to new languages.