论文部分内容阅读
尽管信号子空间方法在语音增强中的应用已经得到了广泛的研究,但是作为制约子空间方法性能的子空间维度估计却一直没有得到较好的解决。针对子空间维度估计问题,本文用多通道语音信号互功率谱矩阵的F范数的统计模型来描述语音信号的先验知识和变化规律,提出了一种基于最大化原则的子空间维度估计方法,在接受原假设的前提下最大化子空间维度。实验证明,在客观语音质量评估和主观测评中,所提算法都取得了更好的结果。与传统方法相比,采用本文方法的多通道语音增强算法可在房间回声、低信噪比等恶劣环境下获得更高的噪声消除和更低的语音畸变。
Although the application of signal subspace method in speech enhancement has been extensively studied, the subspace dimension estimation as a subspace method has not been well solved. In order to solve the problem of subspace dimension estimation, this paper uses the statistical model of F-norm of multi-channel speech signal cross-power spectral matrix to describe the prior knowledge and variation of speech signal, and proposes a subspace dimension estimation method based on maximization principle , Maximizing the subspace dimension while accepting the null hypothesis. Experiments show that in the objective assessment of voice quality and subjective evaluation, the proposed algorithm has achieved better results. Compared with the traditional method, the multi-channel speech enhancement algorithm using this method can obtain higher noise cancellation and lower speech distortion in harsh environments such as room echo and low signal-to-noise ratio.