Acoustic Scene Classification Technique for Active Noise Control

摘要

This paper works on the idea of integrating the acoustic scene classification technique in active noise control (ANC), whereby the major difficulty lies in the limited dataset available from an ANC system in use. A large-scale acoustic scene dataset recorded by a different acoustic device has to be considered. This leads to the research problem of acoustic scene classification by mismatched devices. Hence, an ensemble of the convolutional neural networks (CNNs) is demonstrated with a novel data augmentation method. Monaural samples are processed by the head-related transfer functions (HRTFs) of 24 azimuths to add in the artificial spatial information. Every two symmetrical azimuths are then paired together to provide a tempo-spectral feature consisting of 4 channels. 12 CNNs are trained with respective features. Finally, different ensemble strategies are carried out and their classification accuracies are compared. When the random forest is used as the meta-learner, the ensemble of CNNs achieves the best classification accuracy of 65.1% over 10 acoustic scenes.