Applying K-Nearest Neighbor and Convolutional Neural Network pooling on Mandarin Vowel Recognition

• Main Authors: You-Cheng Lin, 林祐丞
• Other Authors: 李宗寶
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/cgi-bin/gs32/gsweb.cgi/login?o=dnclcdr&s=id=%22107NCHU5337011%22.&searchmode=basic

碩士 === 國立中興大學 === 統計學研究所 === 107 === In the past 10 years, with the rise of the artificial neural network, machine learning has made rapid advance in speech and image recognition, while Chinese pronunciation can be divided into two parts: vowel and consonant. Take ＜ㄈㄚˇ>for example,＜ㄈ>is consonant and ＜ㄚˇ> is vowel . There are 160 types of vowels, 36 types of consonants, and 5 tones which compose 1,391 Chinese pronunciation. In this paper, the k-nearest neighbor(KNN) method and the convolutional neural network(CNN) are used to identify the vowel. The data is recorded from 20 speakers. After sampling, endpoint detection, and frame cutting, the dimension of parameter matrix of each pronunciation data is 53x39, and then attempts to use k-nearest neighbor method , and the convolutional neural network model to identify different hyperparameters with non-pooling, maximum pooling, and average pooling respectively, and explores the effects of various combinations on the accuracy of identification. Because CNN is nonlinear fitting, the recognition rate is much higher than KNN which is linear. The hyperparameter of CNN's highest resolution is 45 frame, kernel size is 5x5, and number of kernel is (512. , 1024, 2048), and using the average pooling, the recognition rate of the vowel can reach 0.9647.