Abstract: There are difficulties in noisy speech recognition, especially low signal-to-noise rations are more difficult. This paper describes briefly six methods for speaker-dependent noisy speech recognition(isolated words). They are LPC prediction error method, one-side auto- correlation sequence LPC, acoustic front end processing, canonical correlation based on compensation method, combination of features method and increase of poles method. The experimental results show that all the six techniques can improve effectively noisy speech recognition, and the best noisy speech recognition rate is above 80%(when SNR=0dB).

Key words: LPC prediction error, one-side autocorrelation sequence LPC, acoustic front end processing, canonical correlation based on compensation, combination of features

摘要： 研究了6种噪声背景下与说话人有关的孤立词语音识别方法。它们是：线性预测误差法，单边自相关线性预测法，语音前端声学处理法，正则相关分析的谱变换补偿方法，特征综合法和同模极点增加法。实验结果表明，这6种方法都有效地提高了噪声环境中语音识别率，其中较好的方法在强噪声环境中(信噪比为0dB)的语音识别率达到80%以上，为信噪比较低的噪声环境中自动语音识别展现了美好前景。

关键词: 线性预测误差, 单边自相关线性预测, 语音前端声学处理, 正则相关分析的谱变换补偿, 特征综合