1、 噪音环境下的语音识别系统的研究 摘 要 语音增强是一个涉及面很广的研究课题，它不仅涉及信号检测、波形估计等传统信号处理理论，而且还与语音特性、人耳感知特性和噪声特性密切相关。因此在语音特性、人耳感知特性及噪声特性的情况下常用语音增强法有：噪声对消法、谱减法、维纳滤波法、卡尔曼滤波法、 FIR 自适应滤波法、基于神经网络的语音增强、基于听觉感知的语音增强、基于小波变换的语音增强方法、自相关法等。 本文主要从以下三种算法进行分析：谱减法 是处理宽带噪声较为传统和有效的方法，利用带噪语音的功率谱估值减去噪声的功率谱估 值实现语音去噪 ； 维纳滤波法是在最小均方准则下实现对语音信号估计的一种滤波器。
2、对于带噪语音信号 ,确定滤波器的冲击响应 ,使得带噪语音信号经过该滤波器后得到最接近于“纯净”的语音信号。 小波变换法应用小波阈值去噪的方法，分析了常用的软、硬阈值函数对语音增强的效果；仿真结果表明小波法效果较明显，而谱减法易产生“音乐噪声”， Wiener滤波增强后的残留噪声类似于白噪声，而不是音乐噪声。 关键词 ： 语音增强 谱减法 小波阈值去噪 维纳滤波法 Research on speech recognition system under noisy environment Abstract Speech enhancement is a broad topic, it not on
3、ly relates to signal detection, Waveform Estimation and other traditional signal processing theory, but also is closely related to the speech characteristics, the human ear and the noise characteristics.Therefore, combined with the speech characteristics and the perceptual properties of human ear an
4、d the noise characteristics.Therefore, in common with speech speech characteristics, human perceptual characteristics and noise characteristics of the case enhancement method:Noise cancellation、 Spectral subtraction、 Wiener filtering method、 Calman filter、 FIR adaptive filtering method、 Speech enhan
5、cement based on Neural Network、Speech enhancement based on auditory perception、 A speech enhancement method based on Wavelet Transform、 The autocorrelation method Etc.This article mainly carries on the analysis from the following three algorithms:Spectral subtraction method is traditional and effect
6、ive broadband noise,Use the noisy speech power spectrum estimation subtract the noise power spectrum estimation implementation of speech denoising.Wiener filtering method is a kind of filter the speech signal estimation in the least mean square criterion implementation.For the speech signal with noi
7、se,Determine the filters impulse response,The noisy speech signal through the filter to obtain the speech signal is closest to the pure.Method of wavelet transform using wavelet threshold denoising,Analysis of the commonly used soft threshold、 effect of hard threshold function for speech enhancement
8、；The simulation results show that the wavelet method has obvious effect, and the spectral subtraction method is easy to produce music noise, Wiener filter enhancement residual noise is similar to white noise, not the music noise. Keywords:Speech enhancement Spectral subtraction Wavelet threshold denoising Wiener filtering method
