Research on the Influence of Multi-GCC Function Feature Fusion on the Accuracy of Time Delay Estimation

The localization of sound sources using the time difference of arrival (TDOA) method heavily relies on accurately estimating the time delay through a microphone array. However, the accuracy of this estimation is often compromised by back-ground noise and reverberation. To address this challenge, this paper introduces a new method for time delay estimation of sound source localization, utilizing a convolutional neural network and a multi-generalized cross-correlation feature fusion (MGCCFF) model. The MGCCFF method constructs a fused feature matrix by employing multiple generalized cross-correlation functions, each possessing distinct time-delay estimation characteristics. This feature matrix is then inputted into a convolutional neural network comprising convolutional and fully connected layers for further processing. The time difference of arrival (TDOA) is subsequently estimated through classification methods. In this study, the proposed MGCCFF-based approach is compared and validated against existing methods through simulations conducted under various signal-to-noise ratio (SNR) and reverberation conditions. The results demonstrate that the MGCCFF method improves TDOA estimation accuracy by approximately 3.36% when compared to existing methods. Besides, the simulation comparison underscores the efficacy of this method, particularly in challenging acoustic environments characterized by background noise and reverberation.