Research On The WFST Based Chinese Speech Recognition Decoder

Speech recognition technology, mainly using the computer speech processingtechnology, to provide a human-machine interface for people, so we can be moreconvenient exchange between man and computer. Since the development of speechrecognition technology has achieved a certain amount of research, many domesticand foreign companies also increase the development and research of large vocab-ulary chinese speech recognition technology. In speech recognition technology, thedecoder is the most critical part. The final search of the speech recognition is com-pleted by the decoder.In recent years, the finite-state transducers are widely used in speech recognitiontechnology. The finite-state transducer not only can be used in analog signal model,but also can simulate many important and complicated grammar structure andgrammar of the natural language characteristics. so it become a powerful tool forspeech research.This article will mainly discuss the application of the finite-state transducer inthe large vocabulary chinese speech recognition system. The basic idea is using thefinite-state transducer to represent the acoustic model, the lexicon and the languagemodel respectively. Then, they will be combined by combining algorithm. And wewill get a complete weighted finite state transducer model. So that we can obtain aspeech recognition research space, which has the same dimension.This thesis is divided into four parts: the first part is the basic concept andtheoretical derivation of the weighted finite-state machine with weight. The secondpart will discuss how to establish the traditional speech recognition models as thefinite-state machines，and to reduce the numbers of states and transition by the”Merge” algorithm. Third, to integrate the weighted finite-state transducers into asearch space using the combining algorithm, and the problem of the optimization.Fourth, we will design and implement the decoder,then we put it to test on the basisof the given test corpus. Finally, we make a comparison between the experimentalresults and the traditional HTK tool based on the recognition result, respectively,recognition rate and recognition speed. Conclusion, to prove the correctness andsuperiority of the recognition system based on the weighted finite-state transducer.