Research On Continuous Speech Recognition Based On A Hybrid HMM/SVM Framework

Most connectionist systems rely on the Hidden Markov Model architecture to model the temporal evolution of speech because of its great flexility of parameter estimation and the multiple choices of its structures and training ways. But the standard Maximum Likelihood estimation criterion for HMM parameter estimation does not guarantee better classification and its output shows the comparability of the samples of the same kind. Support Vector Machine, whose output shows the discrimination of different samples, is based on the principle of structural risk minimization and the control over the generalization offered by Structure Risk Minimum is what makes an SVM a very powerful machine learning technique. That how to make a hybrid SVM/HMM system combine their advantage together is studied in this thesis.Firstly, automatic speech recognition technology and its history and actual state of technology development are introduced, Secondly, the theory of HMM and the structure of the classical HMM-based ASR system are studied. Then, the basic theory of SVM and its successful applications on pattern recognitions are described, and the solutions to problems of placing SVM into speech recognition framework are discussed, such as how to get the fixed training vector and how to convert the SVM output to posterior probability and how to design the classifiers. At last, a hybrid HMM/SVM system is implemented, the testings prove that the hybrid system improves the performance of the original HMM baseline system.