Benchmarked ResNet, DenseNet, MobileNetV2, and a custom six-layer CNN on 6,000+ audio samples across 13 classes. Noise reduction preprocessing on Mel spectrograms improved classification accuracy by 15%.