稀疏重构的压缩感知语声增强模型与算法

Speech Enhancement Model and Algorithm Based on Sparse Signal Reconstruction in Compressive Sensing

语声增强的目的在于消除带噪语声信号中的噪声干扰,提高语声信号的可听度与可懂度。与传统语声增强算法不同,本文利用语声信号与噪声信号的稀疏性差异,提出了一种基于稀疏重构的压缩感知语声增强模型,并导出该模型的数学表达式。基于此语声增强模型,本文还融入了语声信号的稀疏性与非平稳性,提出了语声存在概率为加权因子的加权正交匹配追踪语声增强算法。仿真实验表明本文提出的语声增强模型与算法具有可行性、有效性以及优越性。本算法不仅可以有效的抑制噪声干扰,还可以保留大部分语声信号,达到语声增强的目的。此外,与谱减法和最小均方误差算法比较,虽然本文算法计算量较大,但是其性能优越。

The objective of speech enhancement is to eliminate noise interference in noisy speech signal and to improve both speech quality and speech intelligibility. Different from traditional speech enhancement algorithms, this paper utilizes the difference of sparsity between speech and noise signal, presents the speech enhancement model based on sparse signal reconstruction in compressive sensing and draws its mathematical expression. According to this speech enhancement model, this paper also takes into account the sparsity and non-stationarity of speech signal, and proposes an orthogonal matching pursuit speech enhancement algorithm weighted with speech presence probability. Experimental results show that the pro- posed speech enhancement model and algorithm is feasible, effective and superior. The proposed algorithm not only can e- liminate noise interference but also can reserve most of speech signal. Therefore, the objective of speech enhancement is at- tained. Furthermore, compared with spectral subtraction algorithm and minimum mean square error algorithm, the proposed algorithm is less efficiently computable, however, its performance is better.