Monitoring of Domestic Activities Using Multiple Beamformers and Attention Mechanism

Abstract

Acoustic scene classification is one of the important technologies for classifying domestic activities. When considering domestic activities as acoustic scenes, unlike the general task of acoustic scene classification, there is the problem that the sounds of the target scene and interference scene can become mixed. To deal with this problem, we propose a classification method using multiple beamformers and an attention mechanism. In the proposed method, multiple beamformers for different target directions are prepared and their outputs are input to a classifier. The proposed method then estimates the importance of each beamformer output by using an attention mechanism. To verify the effectiveness of the proposed method, we generated acoustic data by mixing the sounds of the target scene and the interference scene, and conducted a classification experiment. The experimental results confirmed that the F-score could be greatly improved by the proposed method.