HYBRID NEURAL NETWORK MODEL WITH AN ATTENTION MECHANISM FOR SOUND SOURCE COORDINATE ESTIMATION BASED ON TDOA SIGNALS UNDER VARIABLE ACOUSTIC CONDITIONS

Abstract

This paper investigates the problem of sound source coordinate estimation based on Time Difference of Arrival (TDOA) signals using neural network methods. A hybrid neural network model with an attention mechanism is proposed to improve localization accuracy and robustness under noisy environmental conditions. The study considers a system of four microphones arranged in a square configuration, where the input data include TDOA signals and environmental parameters such as temperature, humidity, and wind speed. The training dataset is generated synthetically with consideration of the physical dependence of sound speed on environmental conditions and with additional Gaussian noise to simulate measurement uncertainties. The proposed hybrid model is compared with a classical Multilayer Perceptron (MLP) architecture. Model performance is evaluated using MSE, MAE, RMSE, maximum error, and noise robustness metrics. Experimental results demonstrate that the hybrid model outperforms the classical MLP approach, achieving an 18.4% reduction in MSE, a 7.7% reduction in MAE, a 9.3% reduction in RMSE, a 26.7% decrease in maximum error, and an 18.0% improvement in noise robustness. The obtained results confirm the effectiveness of attention mechanisms for acoustic sound source localization tasks and demonstrate the перспективність proposed approach for practical intelligent localization systems.