基于改进噪声估计的MMSE-LSA语音增强算法

MMSE-LSA SPEECH ENHANCEMENT ALGORITHM BASED ON IMPROVED NOISE ESTIMATION

为解决在非平稳环境下,对数谱最小均方误差估计(MMSE-LSA)算法噪声估计误差较大、语音增强性能受限的问题,提出一种基于改进噪声估计的MMSE-LSA语音增强算法。利用对数能量与谱熵来构建新的语音特征参数能熵比,通过建立能熵比与语音存在概率的正比关系模型,得到语音存在概率估计,再利用平滑后的语音存在概率来更新噪声估计;加入地板阈值来约束MMSE-LSA的谱增益函数,缓解谱增益欠估计引起的语音失真。多种噪声仿真和真实广播信号实验结果表明,算法能较好地跟踪噪声变化,减少语音失真和噪声残留,提高语音信噪比和可懂度。

In the non-stationary noise environment,aimed at the noise estimation error and limited speech enhancement performance of MMSE-LSA algorithm,a MMSE-LSA speech enhancement algorithm based on improved noise estimation is proposed.The logarithmic energy and spectral entropy values were used to construct a new speech feature parameter called energy entropy ratio,and then a relationship model between energy entropy ratio and the probability of speech existence was established,from which the speech existence probability could be calculated.Then the smoothed probability of voice existence was used to update the noise estimate.At the same time,a floor threshold was added to constrain the spectrum gain function of MMSE-LSA,to reduce speech distortion caused by underestimation of spectral gain.A variety of noises were used for simulation experiments,and then verified by real broadcast signals.The experiments prove that the proposed algorithm can better track the change of noise,reduce speech distortion and noise residue,and improve speech signal-to-noise ratio and intelligibility.