利用结构特征的语音压缩感知重建算法

发布时间：2019-03-31 19:14

【摘要】：针对语音信号在变换域中不够稀疏使得压缩感知重建困难的问题,提出了一种利用频域结构特征的重建算法.该算法为单帧语音信号的修正离散余弦变换系数引入幅度和状态2个隐变量,并分别用高斯马尔可夫过程和马尔可夫链对幅度和状态沿频率轴的连续性建模.在此基础上用因子图表示系数及其幅度、状态的联合后验分布,在因子图上用Turbo消息传递迭代求出系数的后验均值,进而重建原始语音信号.与当前几种最新的算法相比,该算法在不同帧长、不同压缩率下均获得更高的重建精度,重建信号在时频图上的能量分布也与原始语音最为接近.可见,利用语音频域系数的连续性,以Turbo消息传递的方式可以在压缩感知中得到较高的重建精度.
[Abstract]:In order to solve the problem that the speech signal is not sparse enough in the transform domain it is difficult to reconstruct the compressed sensing. A reconstruction algorithm based on the structural features of the frequency domain is proposed. In this algorithm, the modified discrete cosine transform coefficients of single frame speech signals are introduced into two hidden variables, amplitude and state, and the continuity of amplitude and state along the frequency axis is modeled by Gaussian Markov process and Markov chain, respectively. On this basis, the coefficients and their amplitudes are represented by the factor graph, and the joint posterior distribution of the states is expressed. The posterior mean value of the coefficients is obtained by means of Turbo message transfer iteration on the factor graph, and the original speech signal is reconstructed. Compared with the latest algorithms, the proposed algorithm achieves higher reconstruction accuracy under different frame length and compression ratio, and the energy distribution of the reconstructed signal on the time-frequency map is the most close to the original speech. It can be seen that by using the continuity of the frequency domain coefficients of speech, we can obtain a higher reconstruction accuracy in compression sensing by means of Turbo message transmission.
【作者单位】：上海交通大学北斗导航与位置服务重点实验室;
【基金】：国家自然科学基金(61171171,61401501) 华为技术有限公司研究基金资助
【分类号】：TN912.3

【相似文献】