Development of an IoT based environmental sound event recognition methods using wireless acoustic sensor networks in smart cities

This study aims to enhance environmental sound classification by evaluating several approaches with a custom dataset recoded with Esp32 device and the publicly accessible ESC-10 dataset. The custom dataset has eight categories: airport, beach, aircraft, highway, lobby, lodge, office, and restaurant, whereas the ESC-10 dataset consists of 400 brief recordings across 10 classes. The first method, used transfer learning with the pre-trained DarkNet53 model, the potential of converting sound recordings into Mel-spectrogram and gray-scale images for feature extraction. Achieving a 97.24% accuracy on the ESC-10 dataset, this model outshined the adaptability and efficiency of pre-trained convolutional neural networks (CNNs) in handling relatively small datasets, reducing computational complexity, and achieving competitive results. The second method showed the advanced neural network architectures, considering Long Short-Term Memory (LSTM) Bidirectional Long Short-Term Memory (Bi-LSTM) and CNN models, on the custom dataset. Bi-LSTM achieved superior performance, showing its super capability for contextual learning and capturing temporal dependencies. These results highlight the importance of flexible model architectures to the specific characteristics of the sound data, with Bi-LSTM achieving the highest training accuracy of 99.4% and test accuracy of 97%. The final method emphasized the application of Mel-Frequency Cepstral Coefficients (MFCC) in conjunction with Support Vector Machines (SVM), concentrating on the impact of feature extraction and segmentation duration. This method achieved a peak test accuracy of 97.39% by employing statistical MFCC features with a 10-second segmentation. The findings underscore the significance of feature extraction and extended segmentation in managing classification jobs and enhancing accuracy. This research greatly enhances the creation of precise and dependable environmental sound classification systems for practical applications through the integration of different methodologies.