Research Of HMM Speech Recognition Technology Based On Improved Training Algorithm

After 50 years'development,speech recognition technology has been widel y applied in our everyday life.Various theories and methods were formed duri ng the development of speech recognition.For example,vector quantization,hi dden Markov models,artificial neural network etc.The contribution of hidden Markov models extends reserching area to large vocabulary,speaker-independen t,continuous speech recognition.However, training algorithm of Hidden Marko v Models is easily trapped in a local optimum.With the deep research of intel ligent algorithm,intelligent speech recognition technology becomes the hot topi c of current research.Speech signal analysis and processing are precondition and basis of speech recognition.First,this thesis discusses speech generating mechanism,speech c haracteristics and the mathematical model of speech,introduces pre-processing methods;then summarizes conventional methods and improved methods;elabora tes vector quantization which is used for speech feature compressing and codin g, and gives steps of LBG algorithm for codebook designing in detail.At the basis of speech signal analysis and processing,this thesis systematically review ed the basic idea of hidden Markov models,forward and backward algorithms,viterbi algorithm and BaumWelch algorithm.On the achievement of modern ar tificial intelligent algorithm,this thesis combine clone selection algorithm and gene cloning technique,and propose an improved BaumWelch algorithm,that Ge ne Cloning BaumWelch algorithm(GCBW algorithm).This algorithm optimizes t he parameter B of hidden Markov models,to solve the problem that BaumWelc h algorithm is easily trapped into a local optimal solution.Experimental results on TIMIT speech corpus show that:Output probability of GCBW algorithm is average 3.67% higher than BW algorithm, speech recog nition rate of test set in state 4 and state 5,GCBW algorithm is higher than BW algorithm,separately 1.49% and 2.64%.Prototype speech recognition syste m is developed on the platform of matlab.Experimental results of specific peo ple online speech recognition further prove that GCBW algorithm is feasible an d effective.It is convenient to analyze model parameters and improve the algo rithm,and it is a good experimental platform for new learners.