DefinePK

Abstract

Environmental sound classification (ESC) is a method to differentiate the audio related to the various environmental sounds. Environmental sounds have a more complex time-frequency structure compared to structured sounds like music and speech. To extract the frequency and time-based features from audio more accurately and effectively, a novel fusion of several features including MFCCs, Mel-spectrogram, spectral skewness, spectral kurtosis and normalized pitch frequency will be evaluated in this study to provide a comprehensive representation of environmental sounds. The fusion will capture various aspects of the input audio data, such as spectral characteristics, statistical properties, and frequency-related information. By using multimodal information fusion, the algorithm will enhance the discriminative power of the model to distinguish between different sounds more effectively. Moreover, the integration of a variety of machine learning models will enhance the robustness and generalization ability of the model. The combination of several machine learning models will reduce the training time and enhance the classification rate of environmental audio under limited computational resources. Furthermore, this thesis will employ three data augmentation methods, namely, time stretch, pitch tuning, and white noise to minimize the probability of overfitting problems due to the limited audios in each class of dataset. This research will evaluate the ensemble model classification accuracy against baseline SVM, RF classifiers, and other state-of-the-art approaches. In UrbanSound8K, ESC-50, and ESC-10 datasets, the highest achieved accuracies using AdaBoost SVM-RF classifiers were ( 94%), (85%), and ( 95%) respectively. The experimental findings demonstrate that the suggested approach achieves superior performance for ESC tasks.