Neural network-based spectrogram feature extraction with KNN ensemble for environmental sound classification

Predicting real-world events is a central objective of machine learning, where pattern recognition in complex signals plays a critical role. Accurate pattern extraction facilitates effective feature representation, which in turn enhances classification performance. In this study, we focus on extracting discrimi-native features using a neural network for environmental sound classification. The network is trained on image repre-sentations of audio signals, primarily spectrograms, which enable the model to capture both temporal and frequency characteristics of sound effectively. The features learned by the neural network are subsequently utilized within a k-nearest neighbors (KNN) ensemble framework for classifica-tion. This ensemble approach allows us to evaluate the ro-bustness and generalization capability of the extracted feature representations. Experimental results demonstrate strong per-formance. The proposed system was evaluated on the DCASE-2017 Acoustic Scene Classification dataset, achiev-ing a classification accuracy of 96.23%. Additionally, testing on the UrbanSound8K dataset yielded an accuracy of 86.7%. These findings indicate that neural network–derived features are highly effective for reliable acoustic scene identi-fication