北京大学学报(自然科学版)

基于单边自相关线性预测噪声中汉语语音识别

黄新宇,吴淑珍   

  1. 北京大学电子学系,北京,100871
  • 收稿日期:1999-10-21 出版日期:2000-09-20 发布日期:2000-09-20

Noisy Chinese Speech Recognition Based on Linear Prediction of One-sided Autocorrelation Sequence

HUANG Xinyu, WU Shuzhen   

  1. Department of Electronics, Peking University, Beijing, 100871
  • Received:1999-10-21 Online:2000-09-20 Published:2000-09-20

摘要: 对含噪语音在自相关域上进行处理,以其自相关函数值为参数进行端点检测,以基于单边自相关序列的LPC倒谱系数作为语音的特征参数进行语音识别,实验表明:这种方法较好地消除了噪声对语音信号的干扰,并获得了较高的识别率。在信号的信噪比低而自相关性又强时,此法能体现出不同一般的优势,为实际应用提供了可能。

关键词: 单边自相关序列, 线性预测编码, 倒谱, 动态时间规正

Abstract: A new way is supplied to increase the recognition rate of Chinese speech in noise. It includes two parts. One is using the autocorrelation function to detect the speech terminal, the other is using the coefficients based on the one-sided autocorrelation sequence to replace the original speech signal and then extract the speech feature to recognize. Isolated word recognition experiment based on DTW shows: it can reduce the disturbance of noise effectively and get the high recognition rate. It is of great advantage to apply when SNR (signal to noise rate) is low.

Key words: one-sided autocorrelation sequence (OSA), linear predictive coding, cepstrum, dynamic time warping

中图分类号: