Data Augmentation Methods for Cross-Device Acoustic Scene Classification

Abstract

Recent advances in deep neural network (DNN) methods have improved the accuracy of acoustic scene classification (ASC). However, these DNN systems have struggled to classify audio scenes across domains, and when faced with domain imbalance in ASC datasets. In this study, we propose an ASC system that addresses these issues using two data augmentation methods. The first method, MixStyleFreq, reduces device mismatch problems by combining the frequency-wise means and standard deviations of convolutional feature maps from different audio scenes. The second method, Spectrum Normalization Augmentation (SpecNormAug), generates additional data for minority devices based on majority devices, improving the representation of minority devices and reducing bias in DNNs toward dominant devices. Our model is built on the efficient MobileNetV2 network, suitable for ASC applications on devices with limited computational capacity. We evaluate our methods on the TAU Urban Acoustic Scene 2020 Mobile dataset, featuring audio scenes recorded by multiple devices. Our approaches significantly improve generalization performance for ASC tasks compared to other data augmentation methods and achieve competitive results compared to state-of-the-art methods.