面向压缩感知语音增强算法的测量矩阵的研究

发布时间：2018-05-09 05:10

本文选题：语音增强 + 压缩感知　；参考：《西北师范大学》2016年硕士论文

【摘要】：近年来,压缩感知技术逐渐成为语音信号处理领域的研究热点,它是对信号采样和数据压缩同时进行的一种新理论,该理论突破了传统的奈奎斯特采样定理的限制。在压缩感知过程中,测量矩阵在信号重构过程中有着相当重要的作用,因此测量矩阵的研究对压缩感知具有重要的理论意义。现阶段压缩感知理论在语音增强领域的研究大多还停留在稀疏表示和信号重构方面,而针对测量矩阵在语音压缩感知技术中的研究还很少。而不同的测量矩阵和重构算法对带噪语音信号的降噪性能有不同的影响。因此,本文以语音增强为目标,结合测量矩阵及其重构算法的特性,研究了测量矩阵和重构算法对语音增强的影响。论文的主要工作如下:1.研究了测量矩阵对语音增强的影响。主要选择了5种常用测量矩阵,分别在不同的噪声环境,不同的信噪比下,进行实验仿真,分析了5种测量矩阵对带噪的单音字、词语、句子增强后的语音信号,并通过两种客观评价方法对去噪后的语音质量进行了语音质量评价。实验结果表明:不管是从去噪信号的波形图还是通过客观评价指标来看,托普利兹矩阵和循环矩阵的去噪效果都略优于其他三种测量矩阵,其中对于白噪声带噪语音,在低信噪比情况下,循环矩阵较优,去噪后的语音平均PESQ值比性能差的伯努利矩阵提高了66.16%。而在高信噪比情况下,比性能相对差的随机高斯矩阵提高了13.54%。2.分别利用了正交匹配追踪(Orthogonal Matching Pursuit,OMP)算法和稀疏自适应匹配追踪(Sparsity Adaptive Matching Pursuit,SAMP)算法研究了重构算法对语音增强的影响。对加入白噪声的带噪语音进行了重构仿真实验,由得出的信噪比结果分析对比两种重构算法的性能优劣。实验结果表明:不管是从重构信号的波形图还是通过客观评价指标来看,传统的OMP算法去噪效果较差,重构语音还包含大量噪声。SAMP算法较之OMP算法,去噪效果较优,它有效地减少了含噪信号中的噪声成分,提高了去噪语音的信噪比,并得出SAMP算法重构语音的信噪比要比OMP算法约高8dB。
[Abstract]:In recent years, compression sensing technology has gradually become a hot topic in the field of speech signal processing. It is a new theory of signal sampling and data compression, which breaks through the limitation of traditional Nyquist sampling theorem. The measurement matrix plays an important role in the process of signal reconstruction in the process of compression sensing, so the study of measurement matrix has important theoretical significance for compression perception. At present, the research of compressed sensing theory in speech enhancement field is still mostly focused on sparse representation and signal reconstruction, but the research on measurement matrix in speech compression sensing technology is still few. Different measurement matrices and reconstruction algorithms have different effects on the noise reduction performance of noisy speech signals. Therefore, the effect of measurement matrix and reconstruction algorithm on speech enhancement is studied based on the characteristics of measurement matrix and its reconstruction algorithm. The main work of the thesis is as follows: 1: 1. The effect of measurement matrix on speech enhancement is studied. Five kinds of commonly used measurement matrices are selected and simulated under different noise environment and different signal-to-noise ratio respectively. The speech signals enhanced by five kinds of measurement matrices are analyzed, including monosyllabic words, words and sentences. The speech quality after denoising is evaluated by two objective evaluation methods. The experimental results show that the denoising effect of the Toplitz matrix and the cyclic matrix is slightly better than that of the other three measurement matrices, both from the waveform diagram of the denoising signal and the objective evaluation index. In the case of low signal-to-noise ratio, the cyclic matrix is better, and the average PESQ value of the de-noised speech is 66.16 times higher than that of the Bernoulli matrix with poor performance. In the case of high signal-to-noise ratio (SNR), the ratio of random Gao Si matrix with relatively poor performance is increased by 13.54.2. The effects of the reconstruction algorithm on speech enhancement are studied using orthogonal Matching pursuit (OMP) algorithm and sparse adaptive matching tracking algorithm (SAMPP). The reconstruction simulation of noisy speech with white noise is carried out, and the performance of the two reconstruction algorithms is analyzed and compared by the result of SNR. The experimental results show that the traditional OMP algorithm has a poor denoising effect, and the reconstructed speech contains a large amount of noise. SAMP algorithm is better than the OMP algorithm in denoising, whether from the waveform diagram of the reconstructed signal or from the objective evaluation index. It can effectively reduce the noise component in the noisy signal and improve the signal-to-noise ratio of the denoised speech. It is concluded that the SNR of the reconstructed speech by SAMP algorithm is about 8 dB higher than that of the OMP algorithm.
【学位授予单位】：西北师范大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TN912.35

【相似文献】