一种基于广义奇异值分解的语音增强算法
A Perceptually Constrained GSVD-Based Approach for Speech Enhancement
-
摘要: 广义奇异值分解的单通道语音增强算法在加性噪声为白噪声的情况下, 效果比较理想.加性噪声为有色噪声的情况下, 通常用一种基于熵奇异值分解 (QSVD) 的方法来处理.对QSVD算法进行衍生, 首先提出了一种基于广义奇异值分解的子空间语音增强算法 (GSVD) .为了处理低信噪比时残留的音乐噪声, 结合人耳的听觉掩蔽效应, 进一步提出了一种基于感官抑制的GSVD (PCGSVD) .试验结果显示, PCGSVD算法能够明显地提高语音质量、可懂度和识别率, 特别是在加性噪声是有色噪声的情况下实验结果明显优于其他的语音增强算法.Abstract: Generalized singular value decomposition is very useful when the additive noise is white.But the residual musical noise is still perceivable under lower signal-to-noise conditions.Therefore, a perceptually constrained GSVD, which is refered to PCGSVD algorithm, is further proposed to incoporate the masking properties of human auditory system to make sure the residual noise to be under the AMTs.
下载:
