Enhancing Laryngeal Cancer Classification Using Voice Data with Octave-band Filters: A Comparative Study of Against Mel-Frequency Filters

Abstract Laryngeal cancer is typically diagnosed through laryngoscopy, but voice-based deep learning algorithms have shown promising performance recently. Most voice-based deep learning algorithms convert voice signals into mel-frequency cepstral coefficients (MFCCs) before processing them through convolutional neural networks (CNNs). In this study, we investigated the use of octave frequency spectrum energy (OFSE), an alternative feature extraction method that improves the representation of low frequency bands in voice signals instead of MFCC. Preliminary results show that using OFSE as input to a CNN improves accuracy by up to 21.6% compared to MFCC. This suggests that OFSE enhances the CNN’s ability to recognize laryngeal cancer-specific features, leading to improved diagnostic accuracy. Additionally, we utilized Class Activation Maps (CAMs) to analyze the important regions for laryngeal cancer diagnosis on the OFSE representation, providing further insights into the effectiveness of OFSE in identifying key diagnostic areas.