基于字典学习的语音增强算法研究

发布时间：2018-05-14 04:24

本文选题：语音增强 + 稀疏表示　；参考：《大连海事大学》2016年硕士论文

【摘要】：在语音通信过程中,语音信号的传递无可避免的会受到噪声的干扰。强噪声背景下会对语音造成严重破坏,不但影响语义辨识降低通信效率,而且长时间的含噪语音还会影响人的听觉和情绪,所以需要在接收端预先加入语音增强环节来改善通信质量。语音增强广泛应用于人机交互系统中,比如人机对话、实时翻译、智能家居等。语音增强作为数字信号处理研究的一个重要方向,其主要作用是削弱噪声,提高语音的清晰度和可懂度。现有的语音增强算法面临的难题是在强噪声背景下提取出的语音仍会在非语音段留下噪声,为语音识别和信息传递工作带来困难。为了解决上述问题,本文主要工作如下：1、对经典语音增强算法中的谱减法进行改进,针对语音活性检测环节,使用短时自相关函数和频谱子带方差设置二级判决阈值,对噪声谱的估计更加准确,实验表明,信噪比和语音质量感知两个评价指标都有所提升。2、字典学习算法研究。从信号表示的角度入手,引入稀疏表示模型,根据稀疏表示后的语音和噪声呈现出的不同特性进行去噪。字典学习算法主要包括两部分,系数追踪算法和字典训练过程。引入的LS-OMP追踪算法对系数更新环节进行了优化,通过实验验证,算法运行速度有所提升；同时对字典参数的最优选择进行讨论。3、对改进后的字典学习算法和经典语音增强算法进行比较总结,实验结果表明,本文提出的算法对非语音段噪声有较好的抑制作用,从而提高了语音质量。
[Abstract]:In the process of speech communication, the transmission of speech signal is inevitably disturbed by noise. Under the strong noise background, it will cause serious damage to the speech, which not only affects the semantic identification and reduces the communication efficiency, but also affects the human hearing and emotion for a long time. In order to improve the communication quality, voice enhancement is needed in the receiver. Speech enhancement is widely used in human-computer interaction systems, such as man-machine dialogue, real-time translation, smart home and so on. Speech enhancement is an important research direction in digital signal processing. Its main function is to weaken noise and improve the clarity and intelligibility of speech. The existing speech enhancement algorithms are faced with the problem that the speech extracted from the strong noise background will still leave noise in the non-speech segment, which makes it difficult for speech recognition and information transmission. In order to solve the above problems, the main work of this paper is as follows: 1. To improve the spectral subtraction in the classical speech enhancement algorithm, the second level decision threshold is set by using the short time autocorrelation function and the variance of the spectrum subband for the detection of speech activity. The estimation of noise spectrum is more accurate. The experimental results show that both SNR and speech quality perception are improved. 2. Dictionary learning algorithm is studied. From the point of view of signal representation, a sparse representation model is introduced, and denoising is carried out according to the different characteristics of speech and noise after sparse representation. Dictionary learning algorithm includes two parts: coefficient tracking algorithm and dictionary training process. The LS-OMP tracking algorithm is introduced to optimize the updating of the coefficients. The experimental results show that the speed of the algorithm is improved. At the same time, the optimal selection of dictionary parameters is discussed, and the improved dictionary learning algorithm and classical speech enhancement algorithm are compared and summarized. The experimental results show that the proposed algorithm has a better suppression effect on non-speech segment noise. Thus, the speech quality is improved.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TN912.3

【相似文献】