Graph-based neural networks with neural ODEs for robust speech processing in environmental and human-centric systems

¹ Turan University, Almaty, Kazakhstan
² Almaty Technological University, Almaty, Kazakhstan
³ International IT University, Almaty, Kazakhstan
⁴ Al Farabi Kazakh National University, Almaty, Kazakhstan

^* Corresponding author: mambetov.saken@gmail.com

Abstract

This paper introduces H-STGNN-ODE-DA, a novel model for voice sentiment analysis that combines multi-scale acoustic feature extraction, hierarchical graph neural networks (GNNs), Neural Ordinary Differential Equations (Neural ODEs), and domain-adversarial adaptation. Designed to enhance accuracy and robustness under real-world conditions, the model was evaluated on IEMOCAP, MELD, and EmoDB datasets, outperforming state-of-the-art approaches such as LSTM, CNN, GCN, GAT, DANN, and SPECTRA. Notably, it achieved a 4.0% improvement over GAT on MELD. Neural ODEs enabled effective modeling of continuous emotional transitions, while domain-adversarial adaptation ensured robustness to domain shifts. Ablation studies confirmed the critical role of each component in achieving high performance. The model demonstrated strong cross-domain transferability, maintaining high accuracy across diverse recording conditions. These results position H- STGNN-ODE-DA as a robust and versatile solution for real-world applications in speech processing, including virtual assistants, social media analysis, and customer service systems.