Determinization of string-to-stringweight finite state transducers in speech recognition

This thesis has carried on a systematic research on the determinization of the string-to-string/weight transducers. It begins from the formal definitions of the string-to-string/weight transducers. Then, three determinization algorithms have been developed. It includes SSW_determinization algorithm only for the complete determinization of the determinizable transducers (AT&T determinization software fsmdeterminize has been successfully reproduced by this algorithm), the partial determinization algorithm PSSW_determinization for the determinization of the non-determinizable transducers, and DSSW_determinization algorithm for the determinization on the demand. These algorithms has been implemented, and a functional test (including ASR test) of these determinization programs has also been carried on.The advantages and disadvantages of these algorithms have been discussed and analyzed. Special attention has been paid to the DSSW_determinization . The DSSW_determinization can be applied to the determinization of both determinizable and non-determinizable transducers, and it has a very low memory cost compared with that of the SSW_determinization and PSSW_determinization