改进的参数自适应的维纳滤波语音增强算法

发布时间：2018-04-22 02:05

本文选题：语音增强 + 参数自适应　；参考：《太原理工大学》2017年硕士论文

【摘要】：近年来,随着信息技术的发展,助听器的技术已日渐成熟,使用智能手机进行语音通信已经成为了人们生活中的必需品,人机交互也逐渐进入了人们的视野,语音已经成为了日常生活中不可或缺的一部分,所以,对语音进行处理成为了迫在眉睫的事情。然而现实生活中,在进行语音处理时,噪声却无处不在,纷繁复杂的噪声成为了处理语音过程中最大的障碍。过去几十年,如何将纯净语音从含噪语音中分离出来一直是各个研究者研究的重点,直至今日,语音增强依然是一个十分活跃的领域。传统的语音增强算法包括谱减法,维纳滤波算法,子空间法,最小均方误差法等。当进行语音增强时,这些传统的算法都没有考虑到不同的噪声对语音信号产生的不同影响。实际上,语音增强过程中包括语音估计和噪声估计两个重要的阶段,在进行噪声估计时,不同的噪声频谱特征不尽相同。本文在此理论基础上提出了根据不同噪声进行参数自适应的语音增强算法。首先,介绍了语音和噪声的相关理论,包括语音和噪声的主要特征,含噪语音的信号模型及噪声评估算法。对近年来的语音增强算法进行归纳总结与比较分析,并对经典的语言增强算法进行仿真实验,分析得出经典算法普遍存在的问题——同一种语音增强算法不能对所有的噪声类型都起到作用。其次,若要使得语音增强算法针对不同的噪声做出不同的处理,就要对噪声进行准确的分类。为了能够对噪声进行准确的分类,本文在对经典的分类算法进行了归纳总结后,阐述了这些分类算法在大数据的处理方面存在的问题,进而提出了本文所使用的分类算法——基于深度置信网络的噪声分类算法(Deep belief networks),并对此算法的分类过程进行了详细的介绍,对其分类的准确性进行了仿真实验并将其与人工神经网络算法的分类准确性进行了对比,结果表明:本文提出的分类算法具有较高的分类准确性。第三,由于基于先验信噪比的维纳滤波语音增强算法能够较大程度的提高语言的质量并且含有较少的音乐噪声,所以本文对基于先验信噪比的维纳滤波语音增强算法进行了详细的推导与分析,并对其进行了相应的仿真实验,得出此算法不能提高所有噪声类型干扰下的语音质量,即而提出了本文的算法——基于噪声分类的参数自适应维纳滤波语音增强算法。本文选用了声音活动检测(VAD)法对噪声功率谱进行估计,并将其与维纳滤波算法结合得到不同噪声条件下的最优的参数组合。将此最优参数组合应用到之前使用深度置信网络算法进行分类后得到的噪声种类中去,并对先验信噪比的高估部分进行修正,最终得到本文提出的语音增强算法。最后,对提出的算法进行实验仿真,将其与经典的基于先验信噪比的维纳滤波算法进行比较,评估算法包括质量评估和可懂度评估,质量评估选用语音质量感知评估(PESQ),可懂度评估选用标准化协方差法(NCM)。在实验的最后对未知的噪声也进行了仿真实验分析,将未知噪声使用本文提出的分类算法匹配到已知噪声中,根据匹配的类型选用最优的参数组合进行语音增强,来证明本文提出的算法具有普遍适用性。结果表明,本文提出的算法能在不降低低信噪比语音可懂度的同时提高语音的质量,并且对未知噪声具有普遍适用性。
[Abstract]:In recent years, with the development of information technology, the technology of hearing aids has matured. The use of smart phone for voice communication has become a necessity in people's life. Human-computer interaction has gradually entered people's field of vision. Voice has become an indispensable part of daily life. Therefore, the processing of voice has become a pressing problem. In the real life, however, in the real life, noise is ubiquitous in the process of speech processing, and the complex noise has become the biggest obstacle in the process of speech processing. In the past few decades, how to separate the pure voice from the noisy speech has been the focus of the researchers. Until today, the speech enhancement is still a one. The traditional speech enhancement algorithms include spectral subtraction, Wiener filtering, subspace, minimum mean square error, etc. when speech enhancement, these traditional algorithms do not take into account the different effects of different noise on speech signals. In fact, speech enhancement includes speech estimation and noise. Two important stages are estimated. In the process of noise estimation, different noise spectrum features are different. Based on this theory, a speech enhancement algorithm based on the adaptive parameters of different noises is proposed. First, the correlation theory of speech and noise, including the main features of sound and noise, and the signal mode of noisy speech, is introduced. Type and noise evaluation algorithms. The speech enhancement algorithms in recent years are summarized and compared, and the classical language enhancement algorithms are simulated and analyzed. The common problems of the classical algorithms are analyzed. The same speech enhancement algorithm can not play a role in all noise types. Secondly, if the speech enhancement is made, the speech enhancement is made. In order to classify the noise accurately in order to make different noises for different noises. In order to classify the noise accurately, after summarizing the classical classification algorithms, this paper expounds the problems existing in the processing of large data by these classification algorithms, and then puts forward the classification calculation used in this paper. Method - Deep belief networks based on deep confidence network, and the classification process of the algorithm is introduced in detail, and the accuracy of the classification is simulated and compared with the classification accuracy of artificial neural network algorithm. The results show that the classification algorithm proposed in this paper is higher. Third, because the Wiener filtering speech enhancement algorithm based on the prior signal-to-noise ratio can greatly improve the quality of the language and contains less music noise, this paper makes a detailed deduction and analysis of the Wiener filtering speech enhancement algorithm based on the prior signal-to-noise ratio, and carries out the corresponding simulation. It is concluded that this algorithm can not improve the speech quality under all noise types. That is, the algorithm, the adaptive Wiener filtering speech enhancement algorithm based on the noise classification, is proposed. The sound activity detection (VAD) method is used to estimate the noise power spectrum, and it is combined with the Wiener filtering algorithm. The optimal combination of parameters under the same noise condition is used to apply the optimal parameter combination to the kind of noise classification obtained before using the depth confidence network algorithm, and to modify the overestimation part of the prior signal to noise ratio. Finally, the speech enhancement algorithm proposed in this paper is obtained. Finally, the proposed algorithm is simulated by experiments. Compared with the classical Wiener filtering algorithm based on the prior signal to noise ratio, the evaluation algorithm includes the quality evaluation and the intelligibility evaluation, the quality evaluation selects the speech quality perception assessment (PESQ), the intelligibility evaluation selects the standardized covariance method (NCM). The simulation experiments are carried out at the end of the experiment. The unknown noise is also analyzed. It is proved that the algorithm proposed in this paper is universally applicable. The results show that the algorithm proposed in this paper can improve the quality of speech at the same time without reducing the speech intelligibility of low signal to noise ratio and to the unknown. Noise is universally applicable.

【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.35

【相似文献】