Standardised speech interfaces - key for objective evaluation of recognition accuracy

Today existing speech recognisers differ not only in their technology from single word to natural language understanding, from hardware based and software-only solutions but also in the recognition accuracy under realtime conditions in the whole variety of environments and networks. Standardised APIs for a defined state diagram of recognition process are the key requirement for integrating different technologies and their evaluation. First time evaluation tests based on standardised interfaces and optimised vocabularies were carried out in an open system environment at Deutsche Telekom’s Research Lab. The approach of a unique lexicon development system enhances the whole process of vocabulary generation. Optimising the phoneme transcription is one of the main future issues at the lab.