Rare sound event detection using superlets and a convolutional TDPANet

Abstract

Rare Sound Event Detection (RSED) focuses on identifying infrequent but significant sound events in audio recordings with precise onset and offset times. It is crucial for applications like surveillance, healthcare, and environmental monitoring. An essential component in RSED systems is extracting effective time-frequency representation as input features. These features capture short, transient acoustic events in an audio input recording, even in noisy and complex environments. Most existing approaches to this RSED problem rely on input features as time-frequency representations, such as the Mel spectrogram, Constant-Q Transform (CQT), and Continuous Wavelet Transform (CWT). However, these approaches often suffer from resolution trade-offs between frequency and time. This trade-off limits their ability to precisely capture the fine-grained details needed to detect these events in complex acoustic environments. To overcome these limitations, we introduce superlets, a novel time-frequency representation that offers super-resolution in both time and frequency domains. To process the high-resolution Superlet features, we have also proposed a Convolutional Temporal Dilated Pyramid Attention Network (TDPANet). This novel neural network architecture incorporates convolutional feature extraction, dilated temporal modeling, multi-scale temporal pooling, and temporal attention mechanisms to enhance event detection accuracy. We evaluate our method on the DCASE 2017 Task 2 rare sound event dataset, which includes isolated sound events and real-world acoustic scenes. Experimental results show that our proposed method significantly outperforms state-of-the-art techniques, achieving an Error Rate (ER) of 0.15 and an F1-score of 92.3%, demonstrating its effectiveness in detecting rare sound events. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2025.