Location-Independent Multi-Channel Acoustic Scene Classification Using Blind Dereverberation, Blind Source Separation, and Model Ensemble
Ryo Tanabe, Takashi Endo, Yuki Nikaido, Kenji Ichige, Nguyen Phong, Yohei Kawaguchi, Koichi Hamada · 2019
This paper presents a location-independent multi-channel acoustic scene classification (ASC) system that avoids spatial overfitting. Generally, ASC suffers from noise and reverberation in real environments. In addition, the ASC performance is decreased by overfitting a dataset, which is the result of learning from acoustic transfer functions enclosed in the dataset. To resolve these problems, we present a location-independent multi-channel ASC system using blind dereverberation, blind sound source separation, pre-trained model-based classifiers, and model ensemble. Experimental results on the DCASE 2018 Task 5 dataset indicate that the proposed system, with an F1 score of 88.4%, outperforms the baseline system. Also, the results indicate that although no one specific function improves the performance dramatically, all functions complement each other through the model ensemble.