ResCapsnet: a capsule network with CRAM and BiGRU for sound event detection

Sound event detection (SED) is a challenging task where ambient sound events are detected from a given audio signal, which includes categorizing the events and estimating their onset and offset times. Deep learning methods such as convolutional neural networks (CNN) and recurrent neural networks (RNN) have achieved promising performance in SED. However, for overlapping sound events, existing deep learning methods are still limited in detecting individual sound events from their mixtures. Inspired by the success of the dynamic routing mechanism of the capsule network (CapsNet), this paper proposes a capsule network model (ResCapsnet-BiGRU) based on a customized residual attention module (CRAM) and bidirectional gated recurrent unit (BiGRU). CRAM is utilized to extract features from log-mel spectrograms that are relevant to sound events. Through dynamic routing, the capsule network can address the overlapping sound events problem. In addition, the BiGRU with time-distributed fully connected layers is adopted to obtain contextual information. Our proposed method was evaluated on two datasets: the Vehicle Weakly Labeled Sound Dataset (VWLSD , DCASE 2017 Task 4) and the Domestic Environment Sound Dataset (DESD , DCASE 2022 Task 4). It achieved F-scores of 62.1% and 75.9% on the Audio Tagging (AT) task, and 54.1% and 59.0% on the sound event detection (SED) task, respectively. The source codes are available at https://github.com/123sunbing/ResCapsnet.git.