Establishment Of Mandarin Large Vocabulary Continuous Speech Recognition Based On Hybrid ANN/HMM Models

Speech recognition is a content-rich widely used technology, which in someapplications is rapidly becoming a key competitive technology. For example, it can beused in automatic speech translation and cross-language communication. This paper ismainly to study a hybrid schema of Artificial Neural Network and Hidden Markov model,applicated to mandarin Large Vocabulary Continuous Speech Recognition System.In this paper, both advantages and disadvantages of ANN and HMM are comparedin detail. And five major categories of hybrid ANN and HMM model are researcheddeeply：early attempts to emulate the algorithm of HMM;ANN to estimate the HMMstate-posterior probabilities；hybrid models based on global optimization；Networks asvector quantizers for discrete HMM; Networks as processor after HMM.With study on advantages of hybrid ANN/HMM, This paper proposes a method ofmandarin large vocabulary continuous speech recognition system based on hybridHidden Markov model and Artificial Neural Network. In the system, a variety of modelswork together. ANN is state model， modeling phonetic properties, and HMMincorporates with linguistics model recognizing utterances. In this way, the system cancombine advantages of both ANN and HMM: HMM has high modeling ability on timesequence structure; ANN has strong ability of nonlinear prediction, modeling, androbustness, as well as facilitation for hardware implementation. At last, in considerationof pronunciation variation with syllable and structural features of ANN, multipathtechnology is added to system and improves modeling of system, with obviousdecrease of deletion errors.The experimental results show that the system effectively take advantages of bothANN and HMM, and improves the performance.