Classification of Non-Speech Sound Signals: An Approach of Machine Learning with MFCC Feature Extraction

The categorization of non-speech sound signals into seven different classes—breath, cough, cry, laugh, sneeze, and yawn—is the subject of this study. The sound signals that were gathered for this research from a variety of sources make up the publicly accessible dataset. In this study, Mel-Frequency Cepstral Coefficient (MFCC) analysis is used to extract sound samples and generates feature vectors with 2 dimensions of sizes 13 x 87. We employed various machine learning (ML) techniques to complete the classification tasks, such as Decision Tree, Random Forest, Logistic Regression, K Neighbors Classifier, Ada Boost Classifier, Gaussian Naive Bayes, and Extra Trees Classifier. A dimensional CNN-based classification method is also used. Evaluating the classification performance in terms of recall, accuracy, precision, F1-Score, and losses forms the basis of the analysis. The results demonstrate that different algorithms perform to varying degrees; the methods with the greatest accuracy and precision ratings are Random Forest and Extra Trees Classifier. To enhance the classification of non-speech sound sources, MFCC feature extraction and ML techniques are using in this research, which helps in healthcare, surveillance, environmental monitoring and detecting criminal activities.