MFCC Feature Extraction and Distance Function Analysis for Smart Animal Sound Classification

Abstract

his study investigates the classification of animal sounds using Mel Frequency Cepstral Coefficients (MFCC) and distance-based functions to develop an efficient and interpretable recognition model. Audio signals were processed using the librosa library in Python to extract sixteen MFCC coefficients representing spectral characteristics of cat and dog sounds. The extracted features were stored in CSV format and analyzed using four distance-based configurations: Mean–Manhattan, Mean–Euclidean, Median–Manhattan, and Median–Euclidean. Each model computed centroid representations of both sound classes and classified samples based on their proximity to these centroids. Experimental implementation was conducted in Google Colab, and results were verified numerically using Microsoft Excel. The findings indicate that the Median–Manhattan model achieved the highest accuracy of 85%, followed by Mean– Manhattan at 80%, Mean–Euclidean at 75%, and Median–Euclidean at 72.5%. The superior performance of the Median–Manhattan model confirms its robustness to outliers and non-uniform feature distributions. This approach offers a computationally lightweight alternative for real-time or embedded applications such as smart animal monitoring, IoT-based acoustic systems, and environmental sound detection.