Sound Event Localization and Detection (SELD) is a problem related to the field of machine listening whose objective is to recognize individual sound events, detect their temporal activity, and estimate their spatial location. Thanks to the emergence of more hard-labeled audio datasets, deep learning techniques have become state-of-the-art solutions. The most common ones are those that implement a convolutional recurrent network (CRNN) having previously transformed the audio signal into multichannel 2D representation. The squeeze-excitation technique can be considered as a convolution enhancement that aims to learn spatial and channel feature maps independently rather than together as standard convolutions do. This is usually achieved by combining some global clustering operators, linear operators and a final calibration between the block input and its learned relationships. This work aims to improve the accuracy results of the baseline CRNN presented in DCASE 2020 Task 3 by adding residual squeeze-excitation (SE) blocks in the convolutional part of the CRNN. The followed procedure involves a grid search of the ratio parameter (used in the linear relationships) of the residual SE block, whereas the hyperparameters of the network remain the same as in the baseline. Experiments show that by simply introducing the residual SE blocks, the results obtained are able to improve the baseline considerably.
View on arXiv