面向双耳助听系统的低复杂度盲源分离算法研究

发布时间：2018-01-11 12:05

本文关键词：面向双耳助听系统的低复杂度盲源分离算法研究　出处：《山东大学》2017年硕士论文　论文类型：学位论文

【摘要】：听损患者选择性注意的能力下降,在多人交谈的场景下确定目标说话者时出现困难,这要求助听器能够帮助患者区别目标声源和干扰声源。盲源分离技术在双耳助听器中用于语音增强时,在消除人声干扰的同时能够通过后处理技术保持信号双耳线索,进而能够帮助患者更好地定位声源,弥补了现有降噪技术的不足,具有很大的应用潜力。卷积混合模型接近于实际环境中的信号混合情况,该模型下有时域和频域两类方法。在回声环境中时域方法收敛性不好,而频域方法通过将时域卷积变为频域相乘获得了高的计算效率。在频域,频域独立成分分析算法(Frequency Domain Blind Source Separation,FDICA)应用广泛,但复杂度相当高,不能满足数字助听器对算法低复杂的要求。针对这个问题,频点选择FDICA通过减少进行ICA迭代的频点数目来降低复杂度。目前的频点选择FDICA均针对小间距麦克风条件下,双耳间距dmic=0.15m条件下还未有相关算法提出,即目前没有能够应用于双耳助听器中的频点选择FDICA。针对现状,本文对双耳间距下频点选择FDICA进行了研究,主要工作如下:1)实现了传统FDICA算法,介绍了复数ICA和解决两种确定性问题等步骤使用的具体算法。ICA算法采用FastICA和SNG ICA的结合算法,讨论了非线性函数和参数的选取问题。置换算法采用基于估计混合矩阵列向量的排序算法。幅度不确定性根据最小失真原则解决。仿真时先通过实验确定ICA的最优参数,该参数设置下大量实验表明实现的传统FDICA具有良好的分离性能。最后简单地估计了传统FDICA每一步的复杂度,发现频点上ICA迭代是主要的复杂度所在,而限制进行ICA的频点数目能够有效地降低这一步的复杂度,这是频点选择FDICA的基本思想。2)在传统FDICA基础上,提出了基于行列式的频点选择FDICA,适用于无回声环境下间距为dmi=0.15m两个麦克风。该算法创新地使用了两次频点选择过程并提出了新的未选频点分离算法。第一阶段频点选择方案采用混合信号协方差矩阵的行列式得到初选频点,第二阶段频点选择方案采用基于箱线图的异常点检测算法对初选频点进一步筛选得到终选频点,两次选择过程保证了选择频点的分离性能。终选频点之外的频点为未选频点,对于这些频点,提出的未选频点分离算法利用估计的相对混合参数构建分离矩阵,同时解决了幅度不确定性问题,简单有效。大量实验表明当阈值设定为0.01时,与传统FDICA相比,提出算法的运行时间下降了约90%,同时分离性能指标均有了显著提升。这说明提出算法能够在降低复杂度的同时提升分离性能。3)在基于行列式的频点选择FDICA基础上,将第一阶段频点选择标准修改为混合信号幅值的互信息,并修改了第一阶段频点选择过程,其他部分不变,提出了基于互信息的频点选择FDICA。通过实验比较了两种频点选择FDICA算法的优劣,发现相比于基于行列式的频点选择FDICA,提出算法虽然复杂度稍有增加,但通过阈值设定所得的最优分离性能有了较大的提升。
[Abstract]:Listen to the ability to damage in patients with selective attention, difficult to determine the target speaker in conversation scenes, which requires hearing aid to help patients to distinguish the target sound source and sound source interference. For blind source separation in binaural hearing aids in speech enhancement, in the elimination of voice interference through postprocessing technique keep the signal of binaural cues, and thus can help patients better sound source localization, makes up the deficiency of the existing noise reduction technology, has great potential in application. The mixed signal of the convolutive mixture model close to the actual environment, there are two kinds of time domain and frequency domain method of the model. In the echo environment in time domain method has better convergence. The frequency domain method by time domain convolution for variable frequency multiplication of high computation efficiency. In the frequency domain, frequency domain independent component analysis algorithm (Frequency Domain Blind Source Separation, F DICA) is widely used, but the complexity is very high, can not meet the digital hearing aid on the low complexity algorithm requirements. To solve this problem, frequency selection by reducing the number of FDICA frequency ICA iteration to reduce complexity. The frequency selection of FDICA for small spacing microphone binaural pitch conditions, under the condition of dmic=0.15m there is no correlation algorithm is proposed, which is currently no frequency point can be used in binaural hearing aids in the selection of FDICA. according to the present situation, this article has conducted the research to binaural spacing frequency selection FDICA, the main work is as follows: 1) the traditional FDICA algorithm, introduces and solve complex ICA algorithm combining specific.ICA algorithm step two a deterministic problem used by FastICA and SNG ICA, and discusses the choice of the nonlinear function and parameter estimation. The replacement algorithm of column sort vector algorithm based on amplitude mixing matrix. According to the uncertainty principle to solve the minimum distortion. The simulation experiments to determine the optimal parameters of ICA, the parameters to achieve a large number of experiments show that the traditional FDICA has a good separation performance. Finally a simple estimation of the traditional FDICA every step of the complexity, found the frequency point ICA iteration complexity is the main. While limiting the number of frequency points of ICA can effectively reduce the complexity of this step, this is the basic idea of.2 frequency point FDICA) on the basis of traditional FDICA, we choose the FDICA frequency determinant based on echo free environment for space dmi=0.15m two microphones. The algorithm of innovation the use of the two frequency selection process and put forward a new non frequency separation algorithm. The determinant of first stage frequency selection scheme based on mixed signal covariance matrix to get the primary frequency, the second phase frequency selection Scheme of the boxplot outlier detection algorithm based on the primary frequency of further screening to obtain the final frequency, two times the selection process to ensure the separation performance of frequency selection. The final frequency frequency outside for frequency, for these frequency, the frequency is not separated by using the method of estimating the relative mixing parameters for constructing the separation matrix, and solves the range uncertainty problem, simple and effective. A large number of experiments show that when the threshold is set to 0.01, compared with the traditional FDICA, the running time of the algorithm is decreased by about 90%, while the separation performance indexes have been significantly improved. This result shows that the proposed algorithm can reduce the complexity and improve the the separation performance of.3) in the choice of FDICA based frequency determinant based on the first stage, the frequency selection criteria for modified mutual information mixed signal amplitude, and modify the first phase frequency selection process, The other part of the same, put forward the frequency point mutual information FDICA. two FDICA algorithms of frequency selection have been compared by experiments based on the FDICA found that compared to frequency determinant based on algorithm complexity, although a slight increase, but the optimal separation performance by setting threshold of income has been greatly improved.

【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.3

【相似文献】