Deep Neural Network for Mandarin and Non-Mandarin Recognition System

• Main Authors: Jayo Wong, 翁嘉佑
• Other Authors: Yuan-Fu Liao
• Format: Others
• Online Access: http://ndltd.ncl.edu.tw/handle/fj2ejc

碩士 === 國立臺北科技大學 === 電子工程系研究所 === 104 === The thesis aims to optimize the Deep Neural Network (DNN) to eliminate the noise to reach better speech recognition of Chinese and non-Chinese. For the building model, the Deep Neural Network has prevailed and had a breakthrough in all aspects in the recent years. However, it is uncertain whether they can be applied to the speech recognition of Chinese and non-Chinese. The thesis transformed the eigenvectors of Universal Background Model (UBM) into low-dimension eigenvectors with i-Vector. The thesis also experimented the effect of Activation Function (Sigmoid, Hyperbolic Tangent, ReLU, ELU) , Dropout, and Training (Back Propagation, AdaGrad) to optimize the Deep Neural Network (DNN). Then, the results were compared with Linear Discriminant Analysis system (LDA) and Probabilistic Linear Discriminant Analysis system (PLDA). Experimental results on the Chinese and non-Chinese speech recognition program evaluation database show the proposed method gained relative performance 0.0879 in EER value and 1.31% in equal error rate (EER) comparing with a baseline provided by program.