Automated bird species recognition using vocalizations is crucial for ecological monitoring and conservation, but faces challenges like high computational costs and a lack of region-specific datasets. While deep learning models are accurate, they are often too resource-heavy for energy-constrained environments. Most existing datasets also focus on North America and Europe, limiting their applicability elsewhere. This study tackles these issues by balancing computational efficiency and accuracy. We evaluate lightweight convolutional neural networks (CNNs) on a new dataset of ten Malaysian bird species. Our results show that MobileNetV3 Small, combined with the STFT time-frequency representation, achieves accuracy within 2.29% of the top-performing model (ResNet50 with Mel spectrogram) while being 46 times more efficient (93 vs. 4,229 MFLOPs). This findings highlight the potential of lightweight models for real-time, energy-efficient biodiversity monitoring, supporting sustainable conservation efforts.