Urbanofonía-3-RNC: una red neuronal convolucional para identificar la taxonomía del paisaje sonoro urbano en espectrogramas generados a partir de audios de ciudades históricas

Urban soundscapes are characterized by the overlapping of multiple sounds, posing challenges for automatic classification via deep learning. This study applies convolutional neural networks (CNNs) with transfer learning to identify diverse sounds in urban environments using visual representations called Mel spectrograms —time-frequency images of audio signals. We created the Urbanphony-3 (UP3) dataset from recordings in historic cities with significant sound overlaps and compared it against two established datasets with minimal overlap: UrbanSound8K (US8K) and Environmental Sound Classification 50 (ESC50). CNNs were trained with each dataset to develop the UP3-CNN, US8K-CNN, and ESC50-CNN models, enabling the automatic recognition of various urban sounds. Model performance was assessed through five-fold cross-validation, using accuracy and loss metrics, as well as confusion matrix analysis and ROC curves. The UP3-CNN model, which classified sounds from environments with frequent overlap, reached an accuracy of 75.2%. In contrast, ESC50-CNN and US8K-CNN, trained with less overlapped sounds, yielded better results (85.6 and 86.3%, respectively). These findings confirm that CNNs have great potential for classifying urban soundscapes, even under natural overlap. However, the performance gap between UP3-CNN and the other models indicates that CNNs are less effective when sounds overlap significantly. Thus, additional strategies are required to improve the results, including data augmentation, transformers, or optimization techniques. Future research should also extend automatic soundscape classification by considering other variables, such as emotional reactions, cultural preferences or contextual influence.