Source identification of weak audio signals using attention based convolutional neural network

Determining the source of a weak sound signal can be difficult, particularly in busy or noisy surroundings. The hand-engineered characteristics and algorithms used in traditional methods for source separation and localization are irresistible to changes in the signal or the noise. In environmental sound classification(ESC) the effectiveness of representative features collected from the environmental sounds is crucial to the classification performance. However, semantically meaningless frames and silent frames frequently hinder the performance of ESC. In order to address this problem, we proposed a new context-aware attention-based neural network for weak environmental sound source identification. Our method is based on the idea that the attention mechanism will be able to concentrate on important elements of the input signal and suppress irrelevant ones, enabling the network to locate the weak sound's source more effectively. To solve the limited capacity problem faced by the attention maps of the attention model we are using context information as additional input. The identification of weak signals and finding corresponding context information is also a challenging problem because of the degradation or noise in the signal. To solve this issue we proposed a novel algorithm based on MFCC for context feature generation. Additionally, the robustness and generalizability of the classification model are improved by using multiple feature extraction techniques which reduces the reliance on any single feature extraction technique. We test our methodology with experiments on datasets of simulated weak sound signals with varying amounts of noise and clutter. We evaluated the performance of our attention-based neural network in comparison to a number of established techniques. Our findings demonstrate that, especially in noisy and congested environments, the proposed model outperforms the baselines in terms of source identification accuracy. Overall, our work illustrates the efficiency of employing an attention-based neural network based on context information for the identification of weak sound sources and implies that this strategy may be a promising approach for further research in this area.