Stacked Ensemble Learning for Environmental Sound Classification Using Bayesian Tensorized Neural Networks

Environmental Sound Classification (ESC) is an important problem in machine learning with real-world applications in environment observation, smart city, and audio surveillance. Hand-engineered feature extraction and traditional classifiers in traditional ESC have difficulty in processing high complexity and unstructure of environment sound. In this work, an intelligent ESC system with a spectrogram feature extraction and a stacked ensemble model is proposed. In the first layer, four algorithms, Linear Regression, Decision Tree, Random Forest, and Support Vector Machine, are blended, and in the second layer, a Bayesian Tensorized Neural Network (BTNN) is adopted for refinement of prediction. The proposed system can effectively extract complex relations in high-dimensional audio feature spaces and enhance accuracy in classification. On three benchmark datasets, such as ESC-10, ESC-50, and UrbanSound8K, experimental comparisons with traditional ML and deep neural networks (DNNs) verify that proposed model outperforms them in accuracy, out-of-class accuracy, and robustness, and present that both ensemble and Bayesian techniques can effectively enhance robustness and generalization for ESC, providing a feasible and efficient model for real-world use cases.