摘要
针对语音识别实际应用过程中的噪声问题,给出了一种新的抗噪声的特征提取算法,即先利用小波变换将语音信号进行小波子带分解,再根据人耳的听觉掩蔽效应,由谱压缩的技术,将小波变换后的子带语音信号进行压缩,从而提取其对应的语音特征。通过MATLAB软件建立实验平台,仿真实验结果表明该语音特征可以在噪声环境下得到较高的识别率。新的特征参数即充分利用了小波的抗噪声特性又有效地降低了语音识别中的训练环境和识别环境间的失配,具有抗噪声的特点。
Aimed at the application of speech recognition, a new method of robust feature extraction is presented. The speech was decomposed by wavelet transformation, and then compressed by spectral compression scheme related to human hearing mask theory. Experimental results of MATLAB simulation show that high recognition rate can be obtained by using of the new feature in noise environment. It can make the best of robust characteristics of wavelet and reduce the difference between training and recognition environment.
出处
《量子电子学报》
CAS
CSCD
北大核心
2009年第4期398-404,共7页
Chinese Journal of Quantum Electronics
关键词
信息处理
语音识别
人工神经网络
谱压缩
information processing
speech recognition
artificial neural networks
spectral compression
