面向混响环境声源定位的 Transformer 特征聚焦建模方法

Traditional sound source localization algorithms in reverberant environments are susceptible to multipath reflections, resulting in insufficient localization accuracy and robustness. To address these issues, a Transformer feature-focused modeling method for sound source localization in reverberant environments is proposed. Using the normalized beamforming power map as input, the proposed method employs the multi-head attention mechanism of the Transformer to model features at different spatial positions, thereby enhancing the global and local feature representations associated with the true source location and suppressing spurious source interference caused by reflections under reverberant conditions. To verify the effectiveness of the proposed method, simulation datasets under different source positions and reverberation conditions are generated using the image source method, and the true source positions are used as supervision labels for training and testing. The results show that, compared with traditional dereverberation methods and early deep learning methods, the proposed method achieves higher localization accuracy and better robustness in reverberant environments, and maintains favorable localization performance especially under highly reverberant conditions. In addition, the proposed method effectively reduces sidelobe levels and weakens the influence of spurious sources. Experiments conducted in an enclosed space further validate the effectiveness of the proposed method in practical scenarios.