Research On The Comparison And Parallelization Of Discriminative Training Of Acoustic Model

Research on the Automatic Speech Recognition has been rapid development inrecent years. More and more voice products are constantly applied in our daily life,all of which are based on the advance of voice technology, especially the research ofthe discriminative training. It can efectively improve the recognition performanceof the system when it is applied to voice continuous speech recognition system. Ithas become a new hotspot in the ASR field in recent years.This article first introduces the development history of speech recognition andthe main stages in the process of development, and then introduces the principalcompositions of the speech recognition system, which we show the detailed intro-duction.Secondly, we introduce the acoustic model based on HMM, and the three basicquestions need to be solved when applied to the speech recognition. We show howto use some algorithms to solve the relevant problems of the speech recognition.Thirdly, we introduce the methods of the discriminative training which is com-monly. At first, we introduced the Maximum Likelihood Estimation (MLE) method,and through the Bayesian Decision Rule to elicit the discriminative training. Byoptimizing the diferent parts, we can obtain the diferent discriminative trainingmethods, e.g. MMIE, MCE, MPE/MWE, and some other the commonly train-ing algorithms. And then we compare the traits of these discriminative trainingmethods. At last, we operate phonetic experiment on HTK platform.Finally, we introduce the parallelization of the discriminate training, and putforward a new parallel algorithm about it. From the last experiment, we find that,the algorithm can reduce the training time efectively without afecting the recogni-tion results when it is combined with discriminate training.