一种改进的子空间语音增强方法

New improved subspace method of speech enhancement

提出一种基于奇异值分解的子空间分解语音增强方法,该方法是利用最小值统计噪声估计法代替传统的VAD方法对噪声进行估计,并利用所得噪声和带噪语音构造的协方差矩阵得到纯净语音的协方差矩阵,并将特征值分解的时域约束和频域约束估计方法推广到奇异值分解方法中,通过奇异值分解、重构得到增强后的语音信息。试验表明:该方法具有较好的去噪效果。

This paper presented a kind of singular value decomposition based on subspace decomposition method of speech enhancement. Using the minimum statistics noise estimation method instead of the traditional VAD method to estimate the noise and use the proceeds of the noise and noisy speech eovariance matrix to be constructed pure speech covariance matrix. Eigen value decomposition of the time-domain constraints and frequency domain constraints estimation method are extended to the singular value decomposition method. Through the singular value decomposition, and reconstruction of post-enhanced voice messages, according to tests show that this method has a good de-noising effect.