Research On Pronunciation Space Modeling Of Non-native Speakers

In this thesis,we research on acoustic modeling for the Xinjiang region of non-native speaking people of Mandarin speech recognition technology.The research of the Xinjiang non-native speaker of Mandarin is not only theoretical meaning,but also has great practical significance.Taking the Uighur speak Mandarin,we attempt to research that how to use a few non-native language speaking people the Mandarin expected data build for non-native speaker of Chinese Mandarin identify.In this paper,the following points:1.Select the best model of the main vowel and initials-finals.As the establishment of standard Mandarin acoustic models,we will use two methods based the main vowel and initials-finals to make the model,and compare the priority of the two models with the experimental comparison.We can get that under the same conditions,the performance of initials-finals is significantly better than the main vowel,and when the number of Gaussian mixture reaches 32,the performance of initials-finals modeling can get the ideal level.2.The method of getting the map rules based on initials-finals.On the problem of dramatic decline in rates when the standard model is applied to Xinjiang non-native speaker,pronunciation dictionary adaptive methods is be used to improve the recognition rate.Compared with Non-native speaker new model which is needed a large number of corpus to be established,we just need to a small amount of non-native speaker corpus on the relative recognition rate increased by 10.44%.3.Access to multi-pronunciation dictionary,to be expanded from the syllable and phoneme level,and select three pruning strategies to make the pronunciation variation pruning compression.By a large number of comparative experiments,we can select the most suitable expansion strategy of Xinjiang non-native speaker pronunciation dictionary.Experimental results show that pronunciation dictionary which is established with the relative maximum pruning strategy based on the vowels is better than others methods and can maximize the improvement of recognition rate.The same time,when the number of dictionary scale is 1.4-1.7 times of the standard dictionary,the recognition rate is in a good level.