分数傅里叶变换域上含噪语音的联合滤波

Combine Filtering in Fractional Fourier Domains for noisy speech

噪声是影响语音识别和说话人识别性能的主要因素,目前常用的降噪方法多是针对平稳噪声的,而针对非平稳噪声的降噪方法很少。而在实际环境中,通常的噪声是非平稳的。本文将含噪语音变换到分数傅立叶域上,提出了一种在分数傅立叶变换域上进行线性最优滤波和中值滤波的联合滤波降噪方法。实验结果表明,该方法对含非平稳噪声的语音的降噪效果明显优于维纳滤波,能够有效地降低非平稳噪声的影响,提高非平稳噪声环境下的语音识别和说话人识别性能。

Many speech enhancement algorithms are implemented to reduce the stationary noise componet embedded in a speech signal. Few algorithms are implemented to reduce the non-stationary noise though the noise in the real environment is non-stationary usually. A new algorithms which combine linearity optimal filtering and median filtering in fractional fourier domains is presented. The experiment showed that it could improve Signal-to-Noise superior to wiener filtering obviously under the coloured noise. It has wide application prospects in speech recognition and speaker recognition.