臺灣博碩士論文加值系統

本文利用卷積神經網路(Convolutional neural network, CNN)來對中文單音進行學習及辨識。主要實驗方向為把單音拆成子音與母音，並在同一模型下預測出子、母音類別，最後組合出單音。其中子音總有36個類別，母音則有160的類別，單音組合則有1391個類別。資料特徵求取方法選用梅爾倒頻譜系數(MFCC)，並以此作為模型輸入數值。本論文將實驗不同卷積層層數、特徵圖(feature map)數和全連接層(full connection layer, FC)的層數、神經元個數對辨識結果的影響。同時地，會探討不同的活化函數(activation function)、初始化方法和BN(batch normalization)、dropout技術的有無是否會影響分類結果。實驗結果發現在使用4層卷積層、3層全連接層並且使用何初始化(He initialization)和BN下獲得最高的單音辨識率。子音、母音和單音辨識正確率分別達到: 96.49%、97.40% 和94.49%。

This thesis is mainly to apply convolutional neural network (CNN) in Chinese monophonic. The main experimental direction is to split the single monophonic into the consonant and the vowel, and predict the consonant and vowel categories under the same model, finally combine them to the monophonic. There are 36 categories of consonants, 160 categories of vowels, and 1391 categories of monophonic. The data feature extraction method uses the Mel-Frequency cepstral coefficients (MFCC) and uses this as the model input value. This thesis will experiment with the effects on the identification results of different convolution layer numbers, feature map numbers, full connection layer (FC) numbers, and number of neurons. At the same time, we will discuss whether different activation functions, initialization methods, batch normalization (BN), and dropout techniques will affect the classification results. The experimental results show that the highest monophonic recognition rate is obtained using four CNN layers, three full connection layers, He initialization and BN. The recognition rates of consonant, vowel and monophonic are: 96.49%, 97.40% and 94.49%, respectively.

目錄
摘要 i
Abstract ii
目錄 iii
附圖目錄 v
表目錄 vi
第一章 緒論 1
1.1研究動機 1
1.2 研究目的 1
1.3相關研究 2
1.3語音辨識介紹 2
1.3.1何謂語音辨識 2
1.3.2語音特性 3
1.3.3語音辨識的應用 3
第二章 語音訊號前處理與特徵參數求取 4
2.1語音訊號 4
2.2資料前處理 5
2.2.1 數位取樣 5
2.2.2 常態化 5
2.2.3 端點偵測 5
2.2.4 切割音框與視窗化 5
2.2.5 預強調 6
2.3特徵參數的求取 6
第三章 研究方法 7
3.1 介紹 7
3.2 感知器 7
3.2.1 活化函數 8
3.2.2 多層感知器 10
3.3 初始化方法 11
3.4 Batch normalization 12
3.5 卷積神經網路 12
3.6 池化層 13
3.7 卷積層結構 14
第四章 實驗與結果 15
4.1 實作軟體 15
4.2 資料來源 15
4.3 主要模型設計 15
4.4 超參數最佳化 16
4.5 實驗結果 17
4.5.1 深度測試 17
4.5.2 寬度測試 18
4.5.3 活化函數測試 19
4.5.4 初始化方法 19
4.5.5 最終模型 20
第五章結論 23
5.1總結 23
5.2改善與建議 24
參考文獻 24
附錄 26

參考文獻

[1] A. Ahad, A. Fayyaz and T. Mehmood, “Speech recognition using multilayer perceptron,” IEEE, pp. 103-109, 2002.
[2] A. Graves, A.-r. Mohamed and G. Hinton, “Speech recognition with deep recurrent neural network,” IEEE International Conference on Acoustics, Speech and Signal Processing, 2013.
[3] Y. Qian, M. Bi, T. Tan and K. Yu, “Very Deep Convolutional Neural Networks for Noise Robust Speech Recognition,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, 12 2016.
[4] 鄒振宏 且 李宗寶, “利用CNN類神經法於中文單音之辨識,” 國立中興大學統計學研究所, 2018.
[5] Kaiming He, Xiangyu Zhang, Xiangyu Zhang, Jian Sun, “Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification,” 2015.
[6] Sergey Ioffe, Christian Szegedy, “Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,” 2015.
[7] 蘇木春 且 張孝德, 機器學習：類神經網路、模糊系統以及基因演算法則。修訂二版, 全華, 2003.
[8] 蘇奕銘 且 李宗寶, 應用MLP、RBF及DNN類神經網路方法於中文母音辨識。碩士論文, 台中: 國立中興大學統計學研究所, 2016.
[9] D.-A. Clevert, T. Unterthiner and S. Hochreiter, “Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs),” 2015.