Whip strike detection using high-sampling-rate audio by evaluating convolutional recurrent neural network configurations and class imbalance strategies

Whip usage in horse racing is regulated in many jurisdictions, with violations typically verified through manual video review—a labor-intensive process unsuitable for real-time enforcement. To automate whip strike detection, this study proposes a convolutional recurrent neural network (CRNN)-based sound event detection (SED) framework designed specifically for impulsive, high-frequency sound events using high-sampling-rate audio data from horse races in Japan. The dataset comprised audio recordings from 24 official horse races with 620 annotated whip strike events. To address the significant class imbalance inherent in the dataset, sample reduction techniques were employed on the majority class. Architectural choices among CRNN models, which combine a convolutional neural network (CNN) for feature extraction and a recurrent neural network (RNN) for temporal analysis, were systematically evaluated, demonstrating that the Convolutional Neural Network 14-Bidirectional Gated Recurrent Unit (CNN14-BiGRU) model achieved the highest F1-score of 69.8%. The results confirm a key insight of this study: capturing impulsive high-frequency sounds is effectively achieved by leveraging high-sampling-rate audio, which provides both accurate high-frequency capture and improved temporal resolution, and by designing CNN architectures to extract richer acoustic features rather than solely enhancing temporal resolution. Furthermore, offline evaluation revealed that the CNN14-BiGRU model achieved detection latencies substantially shorter than real-time audio duration, indicating potential for practical real-time implementation. However, environmental noise and limited dataset size remain significant challenges affecting detection robustness and generalization. This work establishes foundational insights for automatic whip strike detection, guiding future model design and real-time implementations. • CRNN-based system proposed for detecting whip strike sounds in horse racing. • Model performance compared across CNN and RNN architecture combinations. • High-sampling-rate audio improves detection via frequency range and time resolution.