Curriculum Learning to Handle Extreme Class Imbalance for Acoustic Modeling of Forest Elephant Calls

Acoustic monitoring is an extremely useful tool for conservationists to track and identify elephants deep within densely wooded environments. However, traditional acoustic audio event detection approaches achieve limited success for two reasons. First, elephant calls are rare in continuous audio data and vary significantly in their duration, frequency, and harmonics. Second, other rare forest noises can easily be mistaken for elephant calls. To tackle class imbalance, we use majority class undersampling and introduce a data augmentation technique that generates additional realistic positive elephant call training examples. Moreover, to handle rare, challenging background sounds, we introduce a novel curriculum driven training strategy that identifies hard to classify background, and uses such sounds to incrementally improve the model. We show that the curriculum approach significantly improves upon our baseline model. Additionally, we introduce the limit case of the curriculum strategy, captured by a new two-stage modeling framework. Our results show that combining these approaches of data augmentation and curriculum-driven learning leads to significant model improvement, achieving a test F1 score of 0.74 compared to 0.45 for the baseline model.