A CNN-GRU approach to capture time-frequency pattern interdependence for snore sound classification

In this work, we propose an architecture named DualConvGRU Network to overcome the INTERPEECH 2017 ComParE Snoring sub-challenge. In this network, we devise two new models: the Dual Convolutional Layer, which is applied to a spectrogram to extract features; and the Channel Slice Model, which reprocess the extracted features. The first amalgamates an ensemble of information collected from two types of convolutional operations, with differing kernel dimension on the frequency axis and equal dimension on the time axis. Secondly, the dependencies within the convolutional layer channel axes are learnt, by feeding channel slices into a Gated Recurrent Unit (GRU) layer. By taking this approach, convolutional layers can be connected to sequential models without the use of fully connected layers. Compared with other state-of-the-art methods delivered to INTERPEECH 2017 ComParE Snoring sub-challenge, our method ranks 5th on performance of test data. Moreover, we are the only competitor to train a deep learning model solely on the provided training data, except for Baseline. The performance of our model exceeds the baseline too much.