Sound source localization based on residual network and channel attention module

Abstract

This paper presents a sound source localization (SSL) model based on residual network and channel attention mechanism. The method takes the combination of log-Mel spectrogram and generalized cross-correlation phase transform (GCC-PHAT) as the input features, and extracts the time–frequency information by using the residual structure and channel attention mechanism, thus obtaining a better localizing performance. The residual blocks are introduced to extract deeper features, which can stack more layers for high-level features and avoid gradient vanishing or exploding at the same time. The attention mechanism is taken into account for the feature extraction stage in the proposed SSL model, which can focus on the most important information on the input features. We use the signals collected by microphone array to explore the performance of the model under different features, and find the most suitable input features of the proposed method. We compare our method with other models on public dataset. Experience results show a quite substantial improvement of sound source localizing performance.