This paper aims to address the limitations of current deep learning algorithms for sound source localization (SSL), which focus on a single feature and frequency scale, neglecting the integration of multi-scale information. The method developed in this study enhances localization accuracy by effectively using the spatial information and spectral diversity provided by microphone arrays.
The method is based on a multi-scale cross-short-time Fourier transform (STFT) complex-valued convolutional neural network (CCNN). It uses cross-STFT spectra at different scales to capture detailed acoustic information across various frequencies. The effectiveness of the algorithm was validated through both simulations and experimental studies.
Experimental results demonstrate that the proposed multi-scale cross-STFT CCNN not only outperforms the single-scale cross-STFT model but also delivers superior localization performance compared to other advanced methods, achieving consistently higher accuracy. The method shows excellent robustness across various signal-to-noise ratio (SNR) conditions and performs well even on imbalanced datasets, confirming its strong generalization capabilities.
This paper introduces a novel approach to SSL that integrates multi-scale information, addressing a key limitation of existing methods. The findings offer significant value to researchers and practitioners in the field of acoustic signal processing, particularly those focused on deep learning-based localization techniques.
