摘要
提出了一种基于最小统计及短时对数谱幅度最小均方误差估计单通道语音增强算法,并在此基础上对其进行了修正。噪声功率谱估计不需要VAD进行有无语音的检测,并在每帧数据都进行更新,可跟踪变电平噪声。该算法在MMSE-LSA准则下得到谱增益函数,并考虑到纯净声音信号频谱特性,对增益函数进行了修正。实验结果表明,该算法可有效去除噪声,在消除音乐噪声的同时对语音信号产生很小的失真,并易于实时处理。
A modified single channel speech enhancement algorithm is described based on a spectral gain derived from MMSE-LSA and a minimum noise estimation approach. The proposed method does not need a voice activity detector to estimate the noise and renew the noise estimation in every frame to track varying level noise. Considering the properties of clean speech signal, the spectral gain derived from MMSE-ISA is modified. Extensive testing has shown that this algorithm can be implemented in real time and excellent noise suppression is achieved, while avoiding the musical residual noise phenomena and leading weak speech distortion.
出处
《电声技术》
2009年第8期55-59,共5页
Audio Engineering