A Novel Sound Source Localization Method Using Transformer

Sound source localization (SSL) has garnered significant attention in the current era of intelligent systems. However, existing localization methods often suffer from limitations such as poor robustness and low accuracy. This paper proposes a novel SSL method that utilizes the self-attention mechanism in Transformers to model audio features. Specifically, after digitizing the audio signal, we first employ a shallow convolutional neural network (CNN) for feature extraction. Subsequently, the flattened audio features undergo deep feature extraction using a straightforward multi-head self-attention mechanism. Finally, we use a fully connected layer to obtain the predicted direction of arrival (DOA) trajectory. Experiments conducted in various scenarios demonstrate that the proposed method offers more stable and precise localization performance.