Augmented Strategy For Polyphonic Sound Event Detection

Sound event detection is an important issue for many applications like audio content retrieval, intelligent monitoring, and scene-based interaction. The traditional studies on this topic are mainly focusing on identification of single sound event class. However, in real applications, several sound events usually happen concurrently and with different durations. That leads to a new detection task on polyphonic sound event classification along with event time boundaries. In this paper, we propose an augmented strategy for this task, which faces challenges of a large amount of unbalanced and weakly labelled training data. Specifically, the strategy includes data augmentation to enrich training set to eliminate data unbalance, a new loss function that combines cross entropy and F-score, and model fusion to integrate the powers of different classifiers. The performance of the strategy is validated on DCASE2019 dataset, and both the event and segment detections are significantly improved over the baseline system.