低信噪比场景下语音增强算法的研究

发布时间：2018-11-28 16:01

【摘要】：语音作为人们交流和表达情感的一种重要媒介,在日常生活中却总是受到噪声的干扰,因此我们需要对混入背景噪声的干净语音进行语音增强。语音增强算法的最终目标就是对背景噪声进行抑制,改善语音听觉质量,同时保证一定的语音可懂度。人们对语音增强算法的研究已有半个多世纪的历史,这期间涌现过很多经典的语音增强算法,如谱减法、维纳滤波法、幅度谱最小均方误差算法等,且一直为人们所研究。这些算法在高信噪比平稳噪声下,通常可以取得良好的语音增强效果,但是在低信噪比非平稳噪声下,语音增强效果却不尽人意,还有很多需要攻克的难题。所以,在低信噪比非平稳噪声场景下对带噪语音信号进行语音增强仍是当前国内外学者研究的一个热点。本文主要针对对数谱最小均方误差(Log-Spectral Amplitude Minimum Mean-Square Error,LSA-MMSE)算法以及信号子空间算法在低信噪比场景下存在的缺陷提出改进。主要研究工作如下:首先,提出了低信噪比场景下改进的LSA-MMSE算法。针对传统LSA-MMSE算法在强噪声环境下语音信息完整保留效果不佳,本文将Loizou等人提出的大部分语音增强算法对带噪语音进行增强处理后普遍存在两种不同类型失真,这一理论应用到LSA-MMSE算法中。基于这一理论对LSA-MMSE算法提出了改进。以往学者总是将区域Ⅰ的衰减失真和区域Ⅱ小于或等于6.02dB的放大失真所对应的幅度谱归为一类处理,认为这样不会对语音信息的完整保留造成影响,研究表明这样反而会产生更多残留噪声。基于这一点,本文对衰减失真对应的幅度谱、小于等于6.02dB放大失真对应的幅度谱、大于6.02dB放大失真所对应的幅度谱分别采取不同程度的向下约束。另外,低信噪比场景下先验信噪比和增益函数的估计误差对语音增强效果有很大影响,改进的LSA-MMSE算法中分别对它们进行了调整。实验结果表明,低信噪比场景下本文算法更好地保留了语音的主要信息,同时有效抑制了低频部分的背景噪声。其次,提出了低信噪比场景下改进的信号子空间语音增强算法。子空间算法有着良好的去噪效果,但在低信噪比环境下仍然残留较多噪声。本文首先把滤除小于零的特征值及与之对应的特征向量,这一方法应用到传统子空间算法中,以达到优化信号子空间的效果。同时提出使用共享正弦多窗谱的协方差估计方法减小估计误差和计算复杂度。最后对估计的干净语音引入维纳滤波函数进行修正。实验结果表明,在5种常见噪声的低信噪比场景下,改进算法能有效去除背景噪声,改善语音听觉质量,其语音增强效果整体优于改进前的算法。
[Abstract]:As an important medium for people to communicate and express their emotions, speech is always disturbed by noise in daily life. Therefore, we need to enhance the voice of clean speech mixed with background noise. The final goal of speech enhancement algorithm is to suppress background noise, improve the quality of speech hearing, and ensure a certain degree of speech intelligibility. Speech enhancement algorithms have been studied for more than half a century. During this period, many classical speech enhancement algorithms have emerged, such as spectral subtraction, Wiener filter, amplitude spectrum minimum mean square error algorithm and so on. These algorithms can usually achieve good speech enhancement effect under high SNR stationary noise, but in low SNR non-stationary noise, the speech enhancement effect is not satisfactory, and there are still many difficult problems to be solved. Therefore, speech enhancement of noisy speech signal in low SNR non-stationary noise scene is still a hot research topic at home and abroad. This paper focuses on the improvement of the logarithmic spectrum minimum mean square error (Log-Spectral Amplitude Minimum Mean-Square Error,LSA-MMSE) algorithm and the signal subspace algorithm in low SNR scenarios. The main research work is as follows: firstly, an improved LSA-MMSE algorithm in low SNR scenario is proposed. Because the traditional LSA-MMSE algorithm can not preserve the speech information completely in the environment of strong noise, there are two different types of distortion after most of the speech enhancement algorithms proposed by Loizou et al are used to enhance the noisy speech. This theory is applied to LSA-MMSE algorithm. Based on this theory, the LSA-MMSE algorithm is improved. In the past, the attenuation distortion of region I and the amplitudes of region 鈪，

本文编号：2363390