We present findings from a study evaluating audio classification in a Kaggle Multiclass competition. Although the open-source Baseline model was a crucial step, it initially performed poorly with an accuracy of 0.02 public score. The Kaggle platform played a vital role in training and evaluating models to address the complexities of improving a poorly performing model. Rigorous experimentation with ResNet models (18, 34, 50, 101, 152) and the DenseNet models (121, 161, 169, 201) were performed with under-sampling and oversampling to measure the accuracy of a birdcall. The methodology involved identifying peaks, creating audio metadata for training data, training it, and measuring accuracy on the Kaggle platform. The model was tested using a manually collected dataset and cross-validated using the truth table from the Birdwatch website to enhance flexibility. Our approach resulted in competitive results against Google inference, confirming the prevailing literature that downsampling a dataset would lead to the development of adaptability and accuracy across competition datasets.