结点阈值小波包变换语音增强新算法

人耳频率分辨率是非线性的,而小波包算法有灵活的时频分析能力,可较好的模拟人耳基底膜的频率分析特性。本文提出了一种新的基于结点阈值的小波包变换语音增强算法。采用Bark尺度小波包对含噪语音进行分解,在语音信号的子带层次上进行阈值操作,并采用软阈值方法进行阈值处理。采用谱熵法估计结点噪声。实验表明,该算法在多种噪声,尤其是有色噪声和非平稳噪声条件下均有较好的语音增强效果。

New algorithm for speech enhancement using node threshold wavelet packet transform

The frequency resolution of human ear has non-linear characteristics, while wavelet packet transform has flexible time -frequency analysis ability so that it can better simulate the frequency analysis characteristics of human ear. In this paper, a new algorithm is proposed for speech enhancement using node threshold wavelet packet transform method. Bark-scaled wavelet packet is used for noisy speech decomposition, and this method operates threshold at sub-band of speech signal, uses soft threshold to deal with wavelet transform coefficients with different thresholds, and spectral entropy method is applied to estimate the node noise. Experiments show that this algorithm has good speech enhancement effects under various noise conditions, especially under color noise and non-stationary noise conditions.