基于端点检测与压缩感知的语音增强方法  

A Speech Enhancement Method Based on Endpoint Detection and Compressed Sensing

作　　者：倪旭昇 颜上取 赵新民 钱盛友 NI Xu-sheng;YAN Shang-qu;ZHAO Xin-min;QIAN Sheng-you(School of Physics and Electronic,Hunan Normal University,Changsha 410081,China)

机构地区：[1]湖南师范大学物理与电子科学学院,湖南长沙410081

出　　处：《电脑与信息技术》2022年第4期25-29,共5页Computer and Information Technology

基　　金：国家自然科学基金资助项目(项目编号:11774088)。

摘　　要：语音信号在离散余弦变换(DCT)域上具有稀疏性,为压缩感知(CS)的应用创造了条件。将语音端点检测与压缩感知相结合,提出了一种新的语音增强方法。带噪语音通过端点检测得到有话帧与无话帧,利用压缩感知对有话帧进行降噪。将带噪的有话帧在DCT域稀疏表示,采用稀疏循环结构化矩阵进行压缩观测;利用前导无话段得到的噪声估计,提出双迭代正则化正交匹配追踪法(DIMP),通过该算法重构出有话帧;最后在时域叠接有话帧与无话帧,实现语音增强。实验结果表明:在低信噪比下,文章算法的去噪效果优于传统的谱减法和维纳滤波法。The speech signal is sparse on the discrete cosine transform(DCT)domain,which creating conditions for the application of compressed sensing(CS).This paper combines voice activity detection with compressed sensing to propose a new method of speech enhancement.The noisy speech obtains the silence frame and the speech frame through the endpoint detection,and uses the compressed sensing to reduce the noise of the speech frame.The noisy speech frame is sparsely represented in the DCT domain,and the sparse cyclic structured matrix is used for compression observation.Using the noise estimation obtained from the silent preamble frame,a double iterative regularized orthogonal matching pursuit method(DIMP)is proposed,the speech frame is reconstructed by this algorithm.Finally,the speech frame and the silence frame are spliced in the time domain to achieve voice enhancement.The experimental results show that the denoise effect of the algorithm in this paper is better than the traditional spectral subtraction and Wiener filtering methods under low signal-to-noise ratio.

关 键 词：离散余弦变换 压缩感知 语音端点检测 语音增强 双迭代正则化匹配追踪法 

分 类 号：TN391[电子电信—物理电子学] TP391.41[自动化与计算机技术—计算机应用技术]