【作者】 毕云龙；

【导师】 叶以正；

【作者基本信息】 哈尔滨工业大学 ， 微电子学与固体电子学， 2010， 博士

【摘要】 自适应滤波器由于具有在未知环境下良好运行并跟踪输入数据统计量随时间变化的能力,被广泛应用于通信、雷达、声纳以及生物医学工程等领域。在语音通信过程中,声学反馈的存在严重影响语音质量,采用自适应滤波器进行声音反馈抑制是非常有效的方法。本文给出了声音反馈抑制的数学模型,并针对声音反馈抑制中的两个典型应用——远距离声学回声消除和近距离声反馈抑制这两个问题的各自特点,给出解决各自问题的算法设计,并根据近距离声反馈抑制要求实时性强的特点,进行了声反馈抑制算法芯片设计的研究。本文主要研究内容包括以下几个方面:首先,提出远距离声学回声消除算法。在远距离声学回声消除过程中,双方对讲和模型噪声同时存在,严重干扰回声消除算法。目前,单独克服其中一种干扰的研究已经较为成功,而同时克服上述两种干扰的回声消除算法尽我们所知只有一篇文献报道,且该文献将双方对讲的时间限制在2秒钟以内,不能满足实际需要。本文在充分研究现有文献的基础上,对现有算法进行改进,提出了能够同时克服双方对讲与模型噪声干扰的远距离声学回声消除算法——VSS-NLMS-UMDT算法。该算法对双方对讲时间没有长度限制,具有很强的实用价值。其次,提出高收敛精度高收敛速度的最小均方误差(Least Mean Square, LMS)算法。现有文献中报道的LMS类算法均采用瞬时梯度,在对LMS类算法进行改进的过程中都集中在对算法步长的改进。本文在首先构造了物理含义明确的极值为零的目标函数基础上,提出了零极值目标函数系统辨识算法的理论迭代表达式。在实际应用转化过程中采用对算法梯度进行在线求平均值的方式,减小了梯度噪声的影响。算法获得较高收敛精度和收敛速度。此外,算法对输入数据存储量很少,计算过程中数据流规整,并行性高,只包含有加法和乘法操作,便于超大规模集成电路实现,为近距离声反馈抑制算法的提出从理论与实践两方面奠定了基础。最后,通过对零极值目标函数系统辨识算法进行进一步改进,提出声反馈抑制算法,并对声反馈抑制算法进行了VLSI芯片设计,得到芯片的版图设计结果。声反馈抑制算法继承了零极值目标函数系统辨识算法高收敛精度高收敛速度的优点,并且具有更强的并行性。在VLSI芯片设计过程中,在充分挖掘算法内在并行性的基础上,对其中的固定系数乘法转化为移位操作,充分挖掘和利用算法中的数据可重用性。VLSI芯片设计结果显示,所获得的VLSI芯片设计性能完全可以满足实际应用中的吞吐率要求,具有实际应用价值。更多还原

【Abstract】 The ability of an adaptive filter to operate satisfactorily in an unknown environment and track time variations of input statistics makes the adaptive filter widely used in communications, radar, sonar, seismology and biomedical engineering. In the speech communications, acoustic feedback makes bad influence to the quality of voice. The adaptive filter is always used to give solutions to acoustic feedback.The mathematical model of acoustic feedback reduction with adaptive filter is given. And according to the characteristics of long distance acoustic echo cancellation (AEC) and short distance acoustic feedback reduction (AFR), which are the typical application of acoustic feedback, the solution to the problem of their own is proposed. In addition, the very characteristic of AFR is running in real time. To solve the problem, the corresponding circuit implementation is made.The dissertation mainly focuses on the following aspects:Firstly, an AEC algorithm is proposed. During AEC process, the presence of double-talk and under-modeling noise affects the algorithm behavior seriously. At the present time, it’s successful to overcome one of the two noises separately. There is only one paper reported which claimed that it can overcome the two noise at the same time. But the double-talk time is limited in two seconds, which can not meet the requirements in practice. Based on the researching for the reference, an AEC algorithm which can overcome both double-talk and under-modeling noise at the same time is proposed by making modification to the algorithm existed, and the algorithm is called VSS-NLMS-UMDT algorithm. The double-talk time is not limited for the algorithm proposed, which makes it practical.Secondly, one of least mean square (LMS) algorithms with high convergence precision and fast convergence speed is proposed. The LMS algorithms reported all adopt instantaneous gradient, and the modification to LMS algorithm is always made to the step size. A target function with obviously physical meaning and zero-minimum is given first, and a theoretical math expression of zero-minimum target function system identification (ZMTFSI) algorithm is derived. The gradient of the algorithm proposed adopts online-averaging-gradient in practical, which can reduce the gradient noise. The algorithm proposed has high convergence precision and fast convergence speed. In addition, the amount of data storage needed is small, and the data flow is regular during calculation. The algorithm proposed has high parallelism, and includes only summation and multiplication, which makes it very convenient to be realized by VLSI. The ZMTFSI algorithm makes good preparation for AFR algorithm in both theory and practice.Lastly, an AFR algorithm is proposed by modifying the ZMTFSI algorithm, the AFR algorithm is realized by VLSI implementation and the layout of the circuit chip is obtained. The AFR algorithm inherits the characteristics of high convergence precision and high convergence speed in the ZMTFSI algorithm, and has much stronger parallelism than the later one. During the VLSI implementation, the parallelism in the algorithm is mined sufficiently, the reusable data is mined and reused, and translates the multiplication with constant factor into shift operation. The results of VLSI implementation demonstrate that the VLSI implementation can meet the throughput in real application very well, and it’s practical very much.更多还原