Application Of DT And DT-Adaptation In Acoustic Modeling Of ASR

With the rapid development of the Automatic Speech Recognition technology whichis based on the statistical pattern recognition theories, remarkable progress has beenachieved in recent years.Discriminative Training has become one of the standard configuration methodsfor the state-of-the-art acoustic modeling and parameters optimization. Beyond that,these DT criteria have also been adopted to combine with other adaptation strategies,such as Linear Regression (LR) and Maximum A Posterior (MAP). All these methodshave been applied in some real systems, the validity of them has been proved to beuseful by the exciting improvements results. The strategy of DT based linear transformbeing applied in the feature end has also been improved to be useful. All the detail ofthese technologies is described in this thesis.Firstly, this thesis gives an overview and summary on the development history ofASR in chapter one. In the second chapter, the traditional acoustic modeling strategies,such as Maximum Likelihood Estimation(MLE), di?erent Discriminative Training cri-teria(DT) and some updating methods of acoustic model parameters are introducedseperately. The e?ectiveness of DT is obviously presented in our experiments. Wecompared the results on multiple Chinese and English recognition tasks which sharesdi?erent recognition types, the relative performance improvements are all above 15%.Secondly, the traditional adaptation methods are brie?y reviewed in the third chap-ter of this thesis.The new adaptation strategies which combine the discriminative train-ing criteria with the traditional adaptation ones are introduced in detail. They are linearregression with discriminative training(DT-LR) and maximum a posteriori with dis-criminative training(DT-MAP),we expand them with the MWCE criterion creativelyand both of DT-LR and DT-MAP are combined together to be adopted for adaptationin this thesis. The experimental results on both Chinese continuous speech recognitiontasks and English Spelling tasks are compared in detail. From the experimental results,it is obviously that the DT based adaptation strategies outperform the traditional MLEbased ones and they could be new e?ective choice to be applied for adaptation.Finally, the expand application of the discriminative training criteria for the fea-ture end's linear transform method– feature Minimum Phone Error(fMPE), which isstrictly introduced in chapter four. This algorithm is realized based on HTK indepen- dently and it is a useful supplement of the traditional DT criteria for the updating ofacoustic model parameters. From the experimental results, some favorable improve-ments have also been achieved. It is very obviously that the MPE criterion could notonly be useful on the model parameters optimization but also on the feature end's re-finement.