Performance Improvement of Speaker Recognition for Clipped Audio Signals

• Main Authors: Yi-Chun Lin, 林怡君
• Other Authors: 蔡偉和
• Format: Others
• Language: zh-TW
• Published: 2012
• Online Access: http://ndltd.ncl.edu.tw/handle/ysqxnq

碩士 === 國立臺北科技大學 === 電腦與通訊研究所 === 100 === 　　This thesis investigates the problem of speaker verification under the condition that the recorded speech signals are clipped due to the saturation of quantization. The clipping of audio signals is not only unpleasant for human listening but also detrimental for speaker verification systems. Although there are a number of restoration techniques for improving the auditory quality of the clipped speech signals, it is found that the speaker characteristics of the restored clipped speech signals can be significantly changed; hence, the restoration techniques are of little help for speaker verification . To solve this problem, this study proposes improving the speaker verification by pruning the clipped signals instead of restoring them. However, to avoid that the length of a testing speech signal may be shorten severely after the pruning, we develop methods for detecting and discarding the speech frames that contain harmful clipped signals while keeping the speech frames that contain acceptable clipped signals. Our experiments conducted using the NIST2001 SRE database show that the proposed methods can reduce around 10% of the equal error rate of the speaker verification .