Binary Classification of Capuchin Bird Calls via Spectrogram-Enhanced Frequency-Aware Convolutional Neural Networks

Automated recognition of bird vocalizations plays a critical role in ecological research, particularly in challenging environments. In this paper, we propose a frequency-aware Deep Learning (DL) framework for the binary classification of Capuchinbird vocalizations using a tailored Convolutional Neural Network (CNN) and smart spectrogram preprocessing. The model was trained and evaluated using a curated subset of the Z by HP Unlocked Challenge 3-Signal Processing dataset, focusing on short audio clips ranging from 2 to 5 seconds. The preprocessing pipeline included duration standardization, zero-padding, and a novel smart cropping method that emphasizes low-frequency energy concentrations relevant to bird calls. Spectrograms were generated using Short-Time Fourier Transform (STFT) and normalized to enhance biologically informative regions. The CNN achieved outstanding performance, with 99% accuracy, 99.5% precision, 98% recall, and a 98.5% F1-score. Visualization tools, along with confusion matrix analysis, confirmed the robustness, generalization, and minimal overfitting of our model. The results demonstrate the effectiveness of our frequency-aware CNN approach for real-world bioacoustic classification tasks. The ability of the framework to reliably detect rare vocalizations under realistic conditions also makes it a valuable tool for scalable wildlife monitoring.