Digital Filter Impact on Convolutional Neural Network Performance for Environmental Sound Classification

Abstract

Often, telephony-style bandwidth restriction techniques are applied raw to environmental sound classification systems without sufficient validation. To test their effectiveness, this study evaluates the impact of various digital filters (Low-Pass, High-Pass, Band-Pass, Band-Stop) on CNN performance on the ESC-50 dataset. After establishing the Log-Mel Spectrogram as the best input feature (surpassing MFCC), experiments proved that standard Band-Pass filters (300-3400 Hz) and Low-Pass filters actually reduced accuracy. This confirms that environmental sounds require a broad frequency spectrum (broadband), especially at high frequencies. Positive findings were obtained from the use of a low-order High-Pass Filter (HPF) (FIR-32) with a cut-off of 1000 Hz, which successfully increased accuracy to 66.20% above the baseline. Spectral analysis shows that this configuration successfully removes low noise without triggering transient smearing (time distortion). Therefore, this study recommends low-order HPF as the new standard, while suggesting the use of adaptive filters (learnable filters) in the future.