Answers are grounded in this paper and cite their sources.
Environmental sound recognition is an active area of research within the broader field of machine listening, yet most existing studies have primarily focused on speech and music domains. This paper proposes a framework for a home-based machine listening system capable of detecting and classifying non-speech environmental sound events using deep learning techniques. The system leverages a Convolutional Neural Network (CNN) architecture alongside Mel Frequency Cepstral Coefficients (MFCC) for feature extraction, enabling a robust representation of audio signals captured in typical household environments. A dataset composed of 2,000 five-second audio samples, spanning 50 different sound classes commonly found in home settings, was used to train and evaluate the model. These classes include a diverse range of domestic, animal, and other external sounds. The CNN was implemented with a sequential model incorporating convolutional, pooling, and dropout layers to improve generalization and reduce overfitting. The model achieved an average accuracy of 73% using five-fold cross-validation, demonstrating promising performance in classifying real-world environmental audio. The proposed framework has potential applications in intelligent home monitoring, particularly for safety, security, and emergency scenarios where visual systems alone may be insufficient. Additionally, a graphical user interface (GUI) was developed to support live recognition, and monitoring of sound events in real time. Future work will explore multi-audio input, real-time system deployment, and enhanced classification using additional acoustic features.
No takes yet. Share an insight, caveat, or question.
Ricardo A. Catanghal (2025) studied this question.