【摘要】 针对传统声源定位，在噪声和混响较大时声源定位的准确性急剧下降的问题，提出了一种使用二十面体多层分支特征学习网络(eicosahedral multilayer branching feature learning network,简称EMBFLN)结构进行声源定位的方法.首先，在最大可控响应功率波束形成法(steered response power with phase transform,简称SRP-PHAT)的基础上引入最大化及最小化操作，从而得到最小化噪声与混响影响以及最大化真实传输路径信号后的响应功率谱图，并将其作为网络输入送入EMBLFN结构中；然后，将Mish激活函数应用到声源定位的深度学习神经网络中，以此平滑网络的输出并提高模型的泛化能力；最后，通过仿真实验验证所提方法的有效性.此外，为了验证所提模型应用场景的可扩展性，该文还使用了近距离采集的无人机音频数据制作半合成的移动无人机声学场景对模型进行了测试.更多还原

【Abstract】 Aiming at the problem that the accuracy of sound source localization drops sharply when the noise and reverberation are large in the traditional sound source localization method, a method of using eicosahedral multilayer branching feature learning network(EMBLFN) structure is proposed for sound source localization. Firstly, the maximum and minimization operation is introduced on the traditional signal processing method steered response power with phase transform(SRP-PHAT) to obtain the noise and reverberation influence, maximize the response power spectrum after the real transmission path signal, and feed it into the EMBLFN structure as the input of the network. Then, the Mish activation function is applied to the deep learning neural network of sound source localization to smooth the output of the network and improve the generalization ability of the model. Finally, the effectiveness of the proposed method is verified by simulation experiments. In addition, in order to verify the scalability of the proposed model, the model is tested by using the close-range UAV audio data and making a semi-synthetic mobile UAV acoustic scene.更多还原