论文部分内容阅读
低信噪比非稳态噪声环境中的语音增强仍是一个开放且具有挑战性的任务.为了提高传统的基于非负矩阵分解(nonnegative matrix factorization,NMF)的语音增强算法性能,同时考虑到语音信号的时频稀疏特性和非稳态噪声信号的低秩特性,本文提出了一种基于多重约束的非负矩阵分解语音增强算法(multi-constraint nonnegative matrix factorization speech enhancement,MC–NMFSE).在训练阶段,采用干净语音训练数据集和噪声训练数据集分别构建语音字典和噪声字典.在语音增强阶段,在非负矩阵分解目标函数中增加语音分量的稀疏性约束和噪声信号的低秩性约束条件,MC–NMFSE能够更好地从带噪语音中获得语音分量的表示,从而提高语音增强效果.通过实验表明,在大量不同非平稳噪声条件和不同信噪比条件下,与传统的基于NMF的语音增强方法相比,MC–NMFSE能获得较低的语音失真和更好的非稳态噪声抑制能力.
Speech enhancement in low signal-to-noise non-stationary noise environment is still an open and challenging task.In order to improve the performance of traditional non-negative matrix factorization (NMF) -based speech enhancement algorithms, Time-frequency sparse characteristics of signal and low-rank characteristics of unsteady noise signals, a multi-constraint nonnegative matrix factorization speech enhancement (MC-NMFSE) is proposed in this paper. Stage, a speech dictionary and a noise dictionary are constructed by using a clean speech training dataset and a noise training dataset respectively.In the speech enhancement phase, the sparseness constraint of speech components and the low-rank constraint of noise signal are added to the nonnegative matrix factorization objective function , MC-NMFSE can better obtain the representation of the speech component from the noisy speech to improve the speech enhancement effect.Experiments show that under a large number of different non-stationary noise conditions and different signal-to-noise ratio conditions, the MC- MC-NMFSE can achieve lower speech distortion and better unsteady state than the speech enhancement method Sound suppression.