Abstract: Neural network-based sound source localization has received wide attention in recent years.However,it is still challenging to mitigate the problems such as loss of implied DOA location information and small sample data.Therefore,a sound source arrival direction estimation method based on GRU and self-attentive network is proposed.The method uses GRU,which works well for small data sets,as the backbone network to compensate for the difficulty of pure sound data collection.At the same time,it uses sound sources from multichannel recordings to form a training set.After the short-time Fourier transform feature extraction to obtain the Meier spectrogram and acoustic intensity vector,then form the input features superimposed by the multi-channel speech spectrogram and the normalized main feature vector.Avoiding the implicit DOA information corrupted by the combination of speech spectrogram and GCC-PHAT features,effectively mitigating the loss of implicit DOA location information.It is used as input into the convolutional recurrent neural network model for supervised learning to obtain the model parameters.The model output uses 3D Cartesian product coordinate regression to obtain DOA location estimates,and adds a self-attentive network for parameter back-propagation during model training,enables the network to calculate the loss and predict the correlation matrix while training to solve the optimal allocation between predicted and reference localization.Experimental results show that the network has high localization accuracy and robustness under different reverberation conditions and signal-to-noise ratios.

Key words: Sound source direction of arrival estimation, GRU, Convolutional neural network, Recurrent neural networks, Self-attention