Detection and Classification of Acoustic Scenes and Events 2018 Challenge PARTIALLY-SHARED CONVOLUTIONAL NEURAL NETWORK FOR CLASSIFICATION OF MULTI-CHANNEL RECORDED AUDIO SIGNALS Technical Report

This technical paper presents the system used in our submission for task 5 of the DCASE 2018 challenge [1]. We proposed a partiallyshared convolutional neural network, which is a multi-task system that contains a common input (the multi-channel log Mel features) and two output branches, a classification branch, which outputs the predicted class, and a regression branch, which outputs a singlechannel representation of the multi-channel input data. Since the system has a shared network between classification and regression, training for regression is expected to enhance another training for classification and vice versa. Because task 5 aims at classification based on multi-channel audio input, we tried to improve classification performance with this system by training classification and regression together. By applying the proposed system incorporated with parameter tuning of the baseline CNN system, we confirmed that the classification F1 score increased to 89.94% in four-fold cross validation, while the baseline system achieved 84.50% .