Abstract: Aiming at the problem of limited sound source localization model performance caused by the insufficient number of measured data samples, an indoor sound source localization method based on transfer learning is proposed. The method uses a convolutional neural network to build a transfer learning model, pre-trains a large amount of simulated data, and re-trains a small sample of measured data on the basis of the pre-trained model, so as to achieve indoor sound source localization with small sample data. Experiments based on the TAU Spatial Sound Events 2019 dataset show that the transfer learning model can achieve high-accuracy azimuth prediction for different small sample measured data, and the localization performance is better than the traditional convolutional neural network model, has certain value for the application of transfer learning theory in indoor sound source localization.

Key words: Indoor sound source localization, Transfer learning, Small sample, Convolutional neural network