Abstract:

Conventional approaches to acoustic source localization simply based on the signal received by microphone array, are often vulnerable to adverse acoustic conditions with low signal-to-noise ratio or high reverberation. Recent years, the approaches based on pattern recognition and machine learning technology are used to locate source in adverse acoustic environment. A weighted method based on Fisher discriminant theory is introduced to sound source localization based on Fisher Weighted Naive Bayes Classifier（FWNBC）. The eigenvector of each position is calculated by the cross-correlation function weighted by the Phase Transformation（PHAT）. Finally, the source location is estimated by using FWNBC. At the same time, experiments are carried out in a real location system to verify the performance of the improved algorithm. The experimental results show that compared with the Naive Bayes Classifier（NBC）, the FWNBC algorithm effectively improves the accuracy of sound source localization.

Key words: microphone array, GCC-PHAT, Fisher discriminant, Naive Bayes classifier

摘要：

传统的基于麦克风阵列的声源定位方法，往往容易受到低信噪比或高混响等不利的声学条件的影响。近年来，基于模式识别和机器学习技术的方法被用来在恶劣环境下进行声源定位。引入了一种基于Fisher判别理论的加权方法，实现了基于Fisher加权朴素贝叶斯分类器（Fisher Weighted Naive Bayes Classifier，FWNBC）的声源定位。通过基于相位变换（Phase Transformation，PHAT）加权的互相关函数来计算每个位置的特征向量，利用Fisher加权朴素贝叶斯分类器估计声源位置。在实际的定位系统中进行实验，验证改进算法的性能。实验结果表明，与使用朴素贝叶斯分类器（Naive Bayes Classifier，NBC）相比，FWNBC算法有效提高了声源定位的精度。

关键词: 麦克风阵列, GCC-PHAT, Fisher判别, 朴素贝叶斯分类器