The end-to-end framework has been introduced into the binaural localization modeling and achieved higher localization accuracy than the other models, however, the reasonability and interpretability for applying the related neural networks remain unclear. It has been well documented that the auditory system relies on binaural cues for sound localization, and the equalization and cancellation (EC) theory describes how the binaural cues are extracted. In this paper, an end-to-end binaural localization model is proposed based on the EC theory. In the proposed model, a convolution neural network(CNN) with a specifically designed activation function is used to implement the EC theory. The proposed model was trained in synthesized rooms and evaluated in real rooms. Experiment results show that CNN kernels learned by the proposed model are corresponding to binaural cues, and the proposed model outperforms the current end-to-end model by a 10.73% improvement in localization accuracy and a 12.91%improvement in root mean square error(RMSE).
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=21689
(438KB)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
