The Research And Realization Of Mandarin Digit Speech Recognition System Based On Optimum State Number

Mandarin digit speech recognition system has been widely used in different regions in the past decades.However, in real condition, mandarin digit speech recognition system always has quite low accuracy for some digits due to the environment factors such as nose.The thesis has made a series of research on training data, evaluation data and acoustic models.New evaluation speakers are selected for two new categories.Also, through the analysis of the system recognition accuracy, we adjust the state number of the monophone and biphone models of specific digits.Recognition accuracy has been improved to some extent. The main research includes the following:1. We study the structure of mandarin digit speech recognition and learn the algorithm for training and evaluating the parameters, also the recognition process.These help us to know more about the relation between training data and evaluation data in a category, and enlighten us the way to improve the models.2. The thesis finds a new method to select a best group of evaluation speakers for a specific category.For each sets of evaluation speakers, fits a curve to them. We also fit a curve to all the speakers in a category. By measuring the Root Mean Square Error (RMSE) that the evaluation speakers' curve compared to the all speaker curve,we can find a group of evaluation speakers that best represent this category.Using these evaluation speakers, we can evaluation how well we've train our models.3. In the research of improving the accuracy of the system, we analyse the errors of digit 1 and 5, hense, we improve the models by adjust state numbers of monophone and biphone models for these two digits.After the training and evaluation of the new models by adjusting state numbers, we obtain new recognition accuracy with an increasement of 0.60%.