Weakly Supervised Graph Neural Network for Line Spectrum Extraction

Abstract

Mechanically generated sounds, common in industrial process control and surveillance, often exhibit narrowband harmonic features that manifest as line spectra in the time–frequency domain. While convolutional neural networks (CNNs) have been employed for line spectrum extraction, their performance is often hindered by the scarcity of high-quality supervised data. To address this limitation, we explore graph neural networks (GNNs), which explicitly model feature relationships. Among GNNs, graph convolutional networks (GCNs) stand out due to their computational efficiency. In this study, we introduce a GCN model enhanced with a weight tensor to effectively extract line spectral features from graph representations of mechanical sounds. Our approach is tailored for weakly supervised scenarios, where time–frequency masks are noisy and interfere with supervision. By leveraging a tensor product operation, the model projects input graphs into a multi-dimensional embedding space, facilitating the learning of diverse and discriminative representations with minimal computational overhead. Experimental results on audio and underwater acoustic datasets reveal that our method outperforms fully supervised baselines while significantly reducing computational requirements. These results underscore the efficiency and practicality of our framework for real-world acoustic processing applications.