Acoustic scene classification based on generative model of acoustic spatial words for distributed microphone array

In this paper, we propose an acoustic scene classification method for a distributed microphone array based on a combination of spatial information of multiple sound events. In the proposed method, each acoustic scene is characterized by a spatial information representation based on a bag-of-words called the bag of acoustic spatial words. To calculate the bag-of-acoustic spatial words, spatial features extracted from multichannel observations are quantized and then aggregated over a sound clip, that is, each sound clip is regarded as a unit of a"document." Moreover, a supervised generative model relating acoustic scenes and bag-of-acoustic spatial words is also adapted, which enables robust acoustic scene classification. Experimental results using actual environmental sounds show that the proposed approach achieves more effective performance than the conventional acoustic scene classification approach not utilizing a combination of the spatial information of multiple sound events.