Abstract: This paper presents a new method for speech recognition in stationary noise. By using the LPC coefficients of noise to predict all the speech signal, the method gets the LPC prediction error(LPCPE) sequence. Then use it to substitute the speech sequence to detect the speech terminal、extract the speech features and to recognize in a suitable way. Isolated words recognition experiment based on DTW shows: this method is o.k. in automatic detection of speech terminal and extraction of speech feature；and achieves very satisfying recognition rate.

Key words: linear predictive coding (LPC), LPC prediction error, cepstrum, dynamic time warping(DTW)

摘要： 介绍一种平稳噪声环境下语音识别的新的方法。该方法利用噪声的LPC系数去预测语音信号，从而得到LPC预测序列，然后把它代替原语音序列来进行语音端点的检测、语音特征的提取和在合适的匹配方式下的识别。实验结果表明：该法在噪声环境下自动检测语音端点和提取语音信号的特征是可行的，获得了很满意的识别率。

关键词: 线性预测编码(LPC), LPCPE(线性预测误差), 倒谱, 动态时轴弯曲或动态时间规正(DTW)