语音增强技术中的先验信噪比估计算法研究

发布时间：2018-03-27 11:17

本文选题：语音增强　切入点：先验信噪比　出处：《烟台大学》2017年硕士论文

【摘要】：语音是人与人以及人与机器间交流、互通最为便捷的方式之一,但语音信号在传输和处理的过程中,不可避免地要受到外界背景噪声的干扰或损伤,从而导致接受者较难理解和感知,同时也会大大降低语音信号处理系统的整体性能。语音增强的目的在于去除被污染语音中的噪声信号,进而增强语音处理、通信设备的鲁棒性,提高语音的质量和可懂度。本文的研究工作主要围绕语音增强技术中的先验信噪比估计问题开展。众所周知,先验信噪比是语音增强技术中的核心参数,对于其估计的合理性和准确度将在较大程度上影响语音增强系统的总体输出性能。论文在研究了多种先验信噪比估计算法的基础上,详细分析了直接判决算法中平滑因子的取值对于系统整体增强效果的影响,通过设置耦合因子将两个不同性能的直接判决算法进行组合,提出了一种新型的先验信噪比估计算法,并经过理论分析和仿真实验验证了该算法的估计效果。论文的研究和创新工作主要包括:首先,阐述、分析了变换域语音增强算法的特点以及先验信噪比参数对于算法整体输出性能的影响,并对现有多种先验信噪比估计算法的基本原理和优缺点进行了说明和分析。其次,针对最为常用的直接判决先验信噪比估计算法进行了详尽分析和研究,采用对数似然比和幅度分布直方图等客观评判标准研究了直接判决算法下语音增强系统输出信号的失真及音乐噪声残留情况,并根据其实验结果给出了算法总体输出性能随平滑因子变化的规律,进而确定了音乐噪声残留量较低及语音失真程度较小时的平滑因子最优取值。再次,依据对平滑因子选值问题的研究结果,通过引入耦合因子将两个具有不同性能和特点的直接判决算法进行自适应耦合,并基于最小均方误差理论推导获取了耦合因子的最优值,从而提出了一种具有耦合结构的先验信噪比估计新型算法。最后,在多种噪声环境及信噪比条件下对论文提出算法的性能进行了仿真验证,其结果有效表明了提出算法的优良性能。在此基础上对全文进行了总结,并对语音增强技术中先验信噪比估计问题的未来研究方向进行了展望。
[Abstract]:Voice is one of the most convenient ways to communicate and communicate between people and machines, but in the process of transmission and processing, it is inevitable to be disturbed or damaged by background noise. Therefore, it is difficult for the receiver to understand and perceive, and at the same time will greatly reduce the overall performance of the speech signal processing system. The purpose of speech enhancement is to remove the noise signal from the contaminated speech, and then to enhance the speech processing. The research work in this paper focuses on the estimation of prior signal-to-noise ratio in speech enhancement technology. As we all know, prior signal-to-noise ratio (SNR) is the core parameter in speech enhancement technology. The rationality and accuracy of the estimation will greatly affect the overall output performance of the speech enhancement system. The influence of the smoothing factor on the overall enhancement effect of the system is analyzed in detail. Two direct decision algorithms with different performance are combined by setting the coupling factor. A new priori signal-to-noise ratio (SNR) estimation algorithm is proposed, and its effect is verified by theoretical analysis and simulation experiments. The characteristics of speech enhancement algorithm in transform domain and the influence of prior signal-to-noise ratio (SNR) parameters on the overall output performance of the algorithm are analyzed, and the basic principles, advantages and disadvantages of various prior SNR estimation algorithms are explained and analyzed. The most commonly used direct decision prior signal-to-noise ratio estimation algorithm is analyzed and studied in detail. The distortion of output signal and the residual of music noise in speech enhancement system under direct decision algorithm are studied by using logarithmic likelihood ratio and amplitude distribution histogram. According to the experimental results, the rule that the overall output performance of the algorithm varies with the smoothing factor is given, and then the optimal value of the smoothing factor with low residual music noise and low speech distortion is determined. Based on the results of the study on smoothing factor selection, two direct decision algorithms with different performance and characteristics are coupled adaptively by introducing the coupling factor, and the optimal value of the coupling factor is derived based on the minimum mean square error theory. A new transcendental SNR estimation algorithm with coupling structure is proposed. Finally, the performance of the proposed algorithm is verified by simulation under various noise environments and SNR conditions. The results show the excellent performance of the proposed algorithm. Based on this, the paper summarizes the whole paper and looks forward to the future research direction of the prior signal-to-noise ratio estimation in speech enhancement technology.
【学位授予单位】：烟台大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.35

【参考文献】