A New Decoder Design For Large Vocabula

An important problem in large vocabulary speech recognition for agglutinative languages like Turkish is the high out of vocabulary (OOV) rate caused by extensive number of distinct words. Recognition systems using words as the basic lexical elements have difficulty in dealing with such virtually unlimited vocabulary. We propose a new time-synchronous lexical tree decoder design using morphemes as the lexical elements. A key feature of the proposed decoder is the dynamic generation of the lexical tree according to the morphological rules. The architecture emulates word generation in the language and therefore allows very large vocabularies through the defined set of morphemes and morphotactical rules.