Speaker Identification by Empirical Mode Decomposition

• Main Authors: Yi-Huan Lai, 賴亦桓
• Other Authors: Yung-Yaw Chen
• Format: Others
• Language: en_US
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/20500645615568765596

碩士 === 國立臺灣大學 === 電機工程學研究所 === 94 === Timbre is a main feature that one verifies who is speaking. It is the information that is hidden inside the acoustic properties. Using the differences of timbre features in speaker identification has been an open issue over the years. In the literature, most speaker identification systems use LPC-derived Cesptral Coefficients (LPCC) or Mel Frequency Cesptral Coefficients (MFCC) as timbre models. The linear and stationary assumptions of above techniques limit identification performance. In this thesis, we apply an adaptive time-frequency distribution, Hilbert-Huang transform. By decomposing original signal into simple oscillation modes empirically, we can obtain meaningful instantaneous frequencies. These instantaneous frequencies are taken as the input pattern to train the Neural Network classifier. Using these timbre features in the proposed system, we achieve a nice accuracy.