论文部分内容阅读
提出了一种既符合人耳听觉特性又具有良好抗噪性的语音特征分析方法。首先将单边自相关函数序列进行时间方向的平滑处理,提高单边自相关函数的抗噪性,然后用平滑后的单边自相关函数序列代替原信号进行频率规整的LPC分析,最后经倒谱变换得到该特征参数。数字语音识别实验证明:利用该特征参数的语音识别系统的识别性能优于MEL倒谱系数、LPC倒谱系数等传统的语音特征参数。
A speech feature analysis method which is both in line with human ear auditory characteristics and has good noise immunity is proposed. Firstly, the unilateral autocorrelation function sequence is smoothed in the time direction to improve the noise immunity of the unilateral autocorrelation function. Then the LPC analysis of frequency normalization is performed by replacing the original signal with the unidirectional autocorrelation function. Finally, Spectral transformation results in the characteristic parameter. Digital speech recognition experiments show that the recognition performance of speech recognition system using this characteristic parameter is better than the traditional speech characteristic parameters such as MEL cepstrum coefficient and LPC cepstral coefficient.