基于单信道的语音增强算法的研究与改进

发布时间：2018-03-30 23:11

本文选题：语音增强　切入点：谱减法　出处：《广东工业大学》2016年硕士论文

【摘要】：在语音通信中经常受到各种各样噪声的干扰,噪声干扰严重时还可能会使人听觉疲劳或者使语音处理系统性能恶化。为了能够从含有各类噪声的语音信号中尽可能恢复出原始的语音信号,需要使用语音增强技术来抑制或者降低噪声,提高语音信号的质量和可懂度。语音增强按信号获取方式的不同,可以分为单信道、双信道和多信道。目前大多数实际应用的语音信号处理系统多数都还是单信道下工作的,而且由于单信道的语音增强获取得到的信息更少,难度要超过双信道和多信道语音增强。因此,非常有必要对基于单信道的语音增强进行研究和尝试改进。鉴于以上分析,本论文的研究工作主要从如下三个方面进行：1、实际应用环境中噪声利,类多样化,不同的噪声都有着不一样的特性,所以在实际应用中要根据具体的噪声情况选用不同的语音增强算法,才能达到最好的语音增强效果。为此,本文深入阐述了谱减算法、维纳滤波算法和最小均方误差算法的基本原理和实现过程。分别在高斯白噪声、粉红噪声和多说话人babble噪声下进行大量测试。实验结果显示,三种算法在各种噪声环境下均能提高语音的质量,但是不一定都能够提高语音的可懂度。2、对噪声的估计在语音增强技术中至关重要,噪声估计过低时会导致残留较多的背景噪声,而噪声估计过高时会削弱掉微弱的语音信息,不准确的估计最终都会导致语音失真。为此,本文研究了VAD噪声估计算法和基于后验信噪比的时间递归平均的噪声估计算法。由于时间递归平均的算法中存在平滑因子阶跃(0或者1)的现象,本文提出了一种改进措施,改进后保证了平滑因子取值比较合理。最后通过结果显示,改进的算法在低信噪比环境下有比较好的改善效果。3、谱减算法简单高效,实时性很高,适用范围比较广泛,但是谱减算法的缺点是输出均伴有起伏较大且刺耳的音乐噪声。为了解决音乐噪声,人们提出了很多的改进措施。最常用的一种是在减法过程中增加一个或者多个修正系数,以减少音乐噪声的影响,但是修正系数需要根据实验或者经验来确定,适应性比较差。为此,本文研究了基于参数最优控制的谱减算法,并且采用改进的引导判决法来估计先验信噪比,从而确定最优的修正系数。实验结果显示,本文的改进算法对语音的可懂度提高有限,但对语音质量的改善效果明显。
[Abstract]:In speech communication, it is often disturbed by various kinds of noise. Noise interference may also cause hearing fatigue or deterioration of speech processing system performance. In order to be able to recover the original speech signal as much as possible from the speech signal containing all kinds of noise, It is necessary to use speech enhancement technology to suppress or reduce noise and improve the quality and intelligibility of speech signal. At present, most of the speech signal processing systems used in practical applications are still working in single channel, and because of the speech enhancement of single channel, less information can be obtained. It is more difficult than dual-channel and multi-channel speech enhancement. Therefore, it is necessary to study and try to improve the speech enhancement based on single channel. The research work of this paper is mainly carried out in three aspects as follows: 1. In the practical application environment, noise is convenient, various kinds of noise, different noise have different characteristics. Therefore, different speech enhancement algorithms should be selected according to the specific noise in practical application to achieve the best speech enhancement effect. The basic principle and implementation process of Wiener filtering algorithm and minimum mean square error algorithm are tested under Gao Si white noise, pink noise and multi-speaker babble noise respectively. The experimental results show that, The three algorithms can improve the quality of speech in all kinds of noise environments, but not all of them can improve the intelligibility of speech. The estimation of noise is very important in speech enhancement technology. If the noise estimation is too low, it will lead to more background noise, and when the noise estimation is too high, the weak speech information will be weakened, and the inaccurate estimation will eventually lead to speech distortion. In this paper, VAD noise estimation algorithm and time recursive average noise estimation algorithm based on a posteriori signal-to-noise ratio (SNR) are studied. Due to the existence of smoothing factor step 0 or 1) in the time recursive averaging algorithm, an improved method is proposed in this paper. Finally, the result shows that the improved algorithm has better effect of improving in low signal-to-noise ratio environment. The spectral subtraction algorithm is simple and efficient, real-time is very high, and the range of application is wide. However, the disadvantage of spectral subtraction algorithm is that the output is accompanied by large and sharp music noise. In order to solve the music noise, many improvements have been put forward. One of the most commonly used methods is to add one or more correction coefficients in the process of subtraction. In order to reduce the influence of music noise, but the correction coefficient needs to be determined by experiment or experience, the adaptability is poor. Therefore, the spectral subtraction algorithm based on parameter optimal control is studied in this paper. The improved guide decision method is used to estimate the prior signal-to-noise ratio (SNR) and to determine the optimal correction coefficient. The experimental results show that the improved algorithm improves speech intelligibility limited but improves speech quality obviously.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TN912.35

【相似文献】