基于低秩矩阵与稀疏矩阵分解的语音增强方法

发布时间：2019-01-20 12:52

【摘要】：语音通话过程中不可避免受到各种环境噪声的影响,严重影响了语音技术的实际应用。语音增强技术是解决噪声污染的有效途径，但是由于真实环境的复杂性，强噪声和多噪声环境的语音增强仍然没有得到有效解决。在鲁棒主成分分析中（简称RPCA），许多实际的观测量都可以归结为一个低秩分量和稀疏分量相加的模式，借助矩阵的低秩和稀疏分解理论，可以从大噪声或异常值污染数据中恢复原始数据信息。本课题从噪声的低秩本质和语音的稀疏特性出发,并对参数进行约束，提出研究基于低秩与稀疏矩阵分解的语音增强方法，该方法假设噪声时频矩阵具有低秩结构，语音时频矩阵具有稀疏结构，对含噪语音时频矩阵进行正确分解，进而获取有效的语音频谱成分。在本文中，我们提出一个有效的优化算法来解决稀疏矩阵与低质矩阵的分离问题，使得含噪语音的语音增强效果更加明显。实验结果表明,该方法的语音增强方法在相比于其他的语音增强方法中，具有更好的增强效果。基于低秩与稀疏矩阵分解的语音增强方法是一种崭新的语音增强方法，和传统方法在原理和工作方式上有极大不同，该方法具有对各种噪声适应性强、不需要进行端点检测、强噪声环境性能好、以及参数少易调节等优点，有望实现语音增强技术的突破。论文的研究将推动信号处理理论及应用语音技术的发展，具有很高的研究意义和实用价值。
[Abstract]:It is unavoidable to be affected by various environmental noise in the process of voice conversation, which seriously affects the practical application of speech technology. Speech enhancement technology is an effective way to solve noise pollution, but because of the complexity of real environment, the speech enhancement in strong noise and multi-noise environment has not been effectively solved. In robust principal component analysis (RPCA), many actual observations can be reduced to a model of adding low rank component and sparse component, and the theory of low rank and sparse decomposition of matrix is used. Raw data information can be recovered from data contaminated with large noise or outliers. Based on the low rank nature of noise and the sparse characteristic of speech, a speech enhancement method based on low rank and sparse matrix decomposition is proposed, which assumes that the noise time-frequency matrix has a low rank structure. The speech time-frequency matrix has sparse structure, which decomposes the noisy speech time-frequency matrix correctly, and then obtains the effective speech spectrum component. In this paper, we propose an effective optimization algorithm to solve the problem of separating sparse matrix from low quality matrix, which makes the enhancement effect of noisy speech more obvious. Experimental results show that the proposed method has better performance than other speech enhancement methods. The speech enhancement method based on low rank and sparse matrix decomposition is a new speech enhancement method, which is very different from the traditional method in principle and working mode. It has strong adaptability to various kinds of noise and does not need endpoint detection. It is expected to achieve a breakthrough in speech enhancement technology due to the advantages of good performance of strong noise environment and easy adjustment of parameters. The research in this paper will promote the development of signal processing theory and applied speech technology, which has high significance and practical value.
【学位授予单位】：南昌航空大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.35

【相似文献】