Study Of Speech Recognition For Digit Based On HMM And ANN

Speech recognition is a meaningful, application extensive technology, its practicability and interest make people having an urgent application demand to it. Because of speech signals diversity and complexity , current speech recognition efficiency is not high. Efficient identification and development of models and algorithms has became an improtant topic in speech recognition research filed.Firstly, preprocessing, extraction methods of LPC,LPCC and MFCC, which reflect spectrum characteristics of speech signal, are systematically analyzed and researched. And then, two methods of speech recognition are researched: Hidden Markov Model(HMM)and Artificial Neutral Net(ANN). In allusion to disadvantage of classical HMM, DDBHMM and its training and model identification algorithm were respectively discussed. Speech recognition and simulations for mandarin digit based on DDBHMM and HMM are given. It is proved that: DDBHMM compared with conventional HMM can increase discrimination.At last, the defect specifically for classics HMM submits a new model of speech recognition that combined by DDBHMM and Self-organizing Feature Mapping Neural Network(SOFMNN). Speech recognition and simulations for mandarin digit based on CDHMM/SOFM model and DDBHMM/SONN are given. It is proved that: HNN/ANN compared with conventional HMM can increase discrimination and robustness apparently and DDBHMM/SOFMNN compared with CDHMM/SOFM mode can improve discrimination. Have verified a new model of speech recognition superiority , feasibility by brought in this paper...