A novel approach for urban sound classification using visual representations of audio signals presented in this study. Leveraging the UrbanSound8K dataset, audio signals are transformed into visual formats through Chromagram, Short-Time Fourier Transform (STFT), Constant-Q Transform (CQT), and Mel spectrogram methods. These are combined via channel-wise stacking of the three most effective spectrograms to create enhanced visual datasets. Five new datasets were derived from UrbanSound8K to support diverse evaluations. The visual forms of audio data allow for detailed feature extraction and effective input for deep learning models. The study compares classification performance across several architectures, including Visual Geometry Group 19-layer network (VGG19), Visual Geometry Group 16-layer network (VGG16), Residual Network with 50 layers (ResNet50), Mobile Neural Network (MobileNet), Inception Architecture Version 3 (InceptionV3), Densely Connected Convolutional Network with 201 layers (DenseNet201), Neural Architecture Search Network Large (NASNetLarge), Inception combined with Residual Network Version 2 (InceptionResNetV2), and Extreme Inception (Xception). A new model named DeepSoundVisionNet (DSVNet) is proposed, demonstrating superior performance. Using 10-fold cross-validation, DSVNet achieved 95.02% accuracy with stacked spectrograms and 93.56% on Mel spectrograms (batch size 16). STFT yielded 91.15%, CQT 82.29%, and Chromagram 75.93% accuracy. DSVNet shows high capability in handling complex data through visualized audio processing. The research highlights the power of deep learning in smart city applications, environmental sound monitoring, and real-time recognition, offering a foundation for enhancing the precision and efficiency of future sound classification systems.