Blind System Identification for Acoustic Impulse Responses Based on Maximum Likelihood Estimation
Blind System Identification for Acoustic Impulse Responses Based on Maximum Likelihood Estimation
复制标题
DOI：
10.1109/ispacs.2018.8923136
复制
发表时间：
2018-11
期刊：
影响因子：
--
通讯作者：
Saki Ohashi;Hiroki Tanji;T. Murakami
中科院分区：
文献类型：
--
作者：
Saki Ohashi;Hiroki Tanji;T. Murakami
讨论了声脉冲响应的盲系统辨识问题。它是一种仅从观测信号估计未知系统的方法。为了解决盲问题，我们使用单输入双输出模型，其中引入了一个额外的传感器，以获得两个通道的观测信号。为此目的，子空间方法是众所周知的。在这种技术中，使用子信道匹配（SCM）和脉冲响应估计使用的二阶统计量的假设下，观察到的噪声遵循高斯分布。然而，在实践中，真实的环境中的噪声，语音信号遵循非高斯分布。因此，在诸如普通办公室房间的真实的环境中，传统方法的性能恶化。本文提出了一种基于极大似然估计（MLE）的非高斯噪声盲系统辨识方法。特别是对语音信号这类超高斯噪声进行了研究。在我们的方法中，观察到的信号被白化，然后使用对数概率密度函数的导数的近似估计脉冲响应。语音噪声下的仿真结果表明了该方法的有效性。
We discuss blind system identification for acoustic impulse responses. It is an approach to estimating unknown system from only observed signals. To solve the blind problem, we use the single-input double-output model in which an additional sensor is introduced in order to obtain two-channel observed signals. For such a purpose, the subspace method is well known. In this technique, the sub-channel matching (SCM) is used and the impulse response is estimated using the second-order statistics under the assumption that the observed noise follows the Gaussian distribution. In practice, however, the noise in the real environments, e.g., speech signals, follows the non-Gaussian distribution. Therefore, in the real environments such as an ordinary office room, performance of the conventional method deteriorates. In this paper, we propose a method of the blind system identification based on the maximum likelihood estimation (MLE) for the non-Gaussian noise. Especially, we focus on the super-Gaussian noise such as speech signals. In our approach, the observed signals are whitened and then the impulse response is estimated using an approximation of the derivative of the logarithm probability density function. Simulation results using speech noise show the effectiveness of the proposed method.