室内两步法监督式学习双耳声源距离估计

提出一种室内环境下两步法监督式学习双耳声源距离估计算法,该算法通过预先估计声源方位角信息以克服声源方位角的变化对声源距离估计性能的不利影响.该算法第1步利用深度神经网络模型估计声源的方位角,并将不同方位角的双耳信号分类;第2步中对每个方位角的双耳信号采用独立的深度神经网络模型进行声源距离估计,其中距离特征选用双耳信号的一些双耳特征和统计特性。在仿真和实际环境下,本文提出的两步法声源距离估计算法的距离估计准确率比现有算法提高了3%~5%左右,并且在各种不匹配环境下的距离估计准确率比现有算法高出5%~10%左右。实验结果表明利用声源方位角信息可以有效提高双耳声源距离估计算法的性能。 

Two-stage supervised binaural distance estimation in room environments

A two-stage supervised binaural distance estimation,algorithm in room is proposed to overcome the adverse impacts of the change of azimuth angles on the distance estimation performance.At the first stage,the algorithm estimates the azimuth of binaural signals using Deep Neural Networks(DNN) models and classify the binaural signals based on the estimated azimuth.At the second stage,for the binaural signals of each azimuth angle,the distance is estimated by exploiting azimuth-dependent DNN models,in which some binaural features and statistical properties are selected as distance features.In simulated and real acoustic environments,the distance estimation accuracy of the proposed algorithm is about 3%-5% and 5%-10%higher than the existing algorithms in matched and mismatched conditions,respectively.Experimental results demonstrate that the distance estimation performance of binaural distance estimation methods can be effectively improved by utilizing the azimuthal information.