基于小波域谱相减算法的语音增强研究

谱相减算法是目前常用的语音增强算法，它常在处理宽带噪声中使用，具有运算量小、效果明显等特点，但该算法是基于短时傅立叶变换（STFT）基础上的，是一种单分辨率的信号分析方法。与STFT本质不同的小波变换是一种变分辨率的时频联合分析方法，当分析低频信号时，其时间窗口很大，而当分析高频时，其时间窗口减小，这恰恰符合实际中对非平稳信号进行分析的自然规律。基于小波域的谱相减算法，是将带噪语音信号进行Mallat多尺度分解，然后分别对各尺度下的信号进行谱相减运算，再逐一进行小波重构，得到去噪后的语音信号。仿真结果表明。该方法不但有效地提高了语音信号的信噪比，而且也在很大程度上改善了语音的失真程度，不失为一种有效的语音增强算法。

Research of Speech Enhancement Based on Wavelet-Domain Spectral Subtraction Algorithm

The spectral subtraction algorithm is now the speech enhancement algorithm of general application. It is widely used in processing of broadband noises, which has the characteristics of low operational quantity and high effect. Being based on STFT, it is a kind of method of single resolution capacity. Deffering from STFT in nature ,the wavelet transform is a multi-resolution analysis method with time and frequency. When analyzing signals of low frequencies, the time window is large and when analyzing high frequencies, the time window is small. This is suitable for the law of practical non-steady signal analysis. The algorithm of spectral subtraction based on the wavelet domain proceeds by decomposing a noisy signal with the Mallat multi-scale way, taking subtraction operation for the signal of each measure scale and reconstructing wavelets one by one. Then the denoised speech signal is obtained. Smulation result shows that this method can increase the SNRs of speech signals effectively and improve the distortion level of voices greatly. It is an effective algorithm of speech enhancement.