The Study of Application to the Speaker Recognition by Combining Support Vector Machine with Decision Tree

• Main Authors: Chun-Chieh Lee, 李俊頡
• Other Authors: 陳國在
• Format: Others
• Language: zh-TW
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/20109655192830535507

碩士 === 國立臺灣大學 === 工程科學及海洋工程學研究所 === 101 === In recent years, the rapidly developing devices including smart phones, tablet PCs and cloud network spread around our life, however the language is still the main way of communication for human being. Because of highly developing of sciences and technologies, the speech will become the main medium of communication between the technical products and the human being in the near future, and for safety protection the voiceprint will be combined with the previous text passwords. Therefore, the enhancement on the identification accuracy must be the foundation for any application. This study is to focus on the topics dealing with the speaker verification to make further experimental discussion, in which the proposal of new identification progress will make the system performance be enhanced. This structure basically includes the most widely used Mel-scale Frequency Cepstrum Coefficients as a speech feature parameter, and further using the combination of the decision tree with the idea of support vector machine to cause two classification ways by using various types of grouping to form a two-phase classification of integration application. Accordingly, this study is to propose integrated verification system by combining the two-phase Classification with the support vector machine and decision tree. Approving through the results of experiment shows that the system as selected can obtain the improvement in its recognition rate. As studied the classification accuracy can be improved by 2.501% compared with the basic support vector machine classification method. While under disturbance by noise it is possible to get high recognition rate only when Multi-function Training mode is added.