Data Driven Mel Filter Bank Design for Environmental Sound Analysis

Rothstein, Michael

Etd

Data Driven Mel Filter Bank Design for Environmental Sound Analysis

Público Deposited

Audio classification is a vital technique in environmental monitoring, facilitating the automatic categorization of audio data into predefined classes based on acoustic features. From identifying wildlife vocalizations to assessing urban noise pollution levels, its applications are diverse and pivotal in understanding and managing ecosystems and urban environments. The conventional audio classification method often utilizes Mel Frequency Cepstral Coefficients (MFCC) extracted from audio files as input to a Deep Neural Network (DNN) classifier. However, its effectiveness is limited by a fixed filterbank structure, designed for the human audio range but lacking optimization and adaptability to diverse datasets. To address this, we propose a customized MFCC approach (Pertinant Spectral Characteristic MFCC), aligning the filterbank with dataset-specific frequency power distribution peaks, thus enhancing classification accuracy and adaptability. Through a comparative analysis across various environmental datasets, including ESC50, UrbanSound8K, and Gunshot our study demonstrates the superiority of the Pertinant Spectral Characteristic MFCC (PSC-MFCC) approach. Specifically, we observed a notable 4.5% increase in classification accuracy and a 1.47% decrease in standard deviation compared to the traditional MFCC method, showcasing its potential to significantly enhance audio classification accuracy and precision. These findings underscore the practical utility and efficacy of the proposed methodology in environmental audio classification tasks. By accurately capturing and distinguishing features within diverse frequency ranges across classes, the PSC-MFCC approach offers a promising avenue for advancing audio classification techniques in environmental monitoring and conservation efforts.

Creator