<![CDATA[Individuals with hearing impairments often have difficulty recognizing environmental sounds that are important for daily activities. This study presents an Android-based environmental sound classification application built on a Convolutional Neural Network (CNN) using the ResNet-50 architecture. The model was trained on the ESC-50 dataset (2,000 samples across 50 classes); each audio file was converted into three-channel images (log-Mel spectrogram, delta, delta-delta) as model input. Hyperparameter tuning identified an optimal configuration (epoch=100, batch size=16, learning rate=0.001). The trained model achieved strong performance with training accuracy ≈ 93.19% and testing accuracy ≈ 92%, and average precision ≈ 0.93, recall ≈ 0.92, and F1-score ≈ 0.92. Field tests revealed degraded performance under high noise levels and at increased distances; usability evaluation yielded Usefulness = 90.30%, Satisfaction = 87.57%, and Ease of Use = 89.09% (mean = 88.98%). These results indicate ResNet-50 is effective for environmental sound classification in controlled settings, while enhanced pre-processing (noise handling) is recommended for robust real-world deployment.]]>
Copyrights © 2026
