Recognizing emotions in speech is challenging, especially for low-resource languages like Urdu. Existing methods often struggle to capture subtle emotional nuances due to limited data. This study addresses this issue by exploring the synergy between Siamese Networks and Convolutional Neural Networks (CNNs) for Urdu speech emotion recognition. Using the SEMOUR + dataset with Mel-Frequency Cepstral Coefficients (MFCCs) as features, the Siamese model demonstrated high training accuracy but faced challenges in generalization. In contrast, a CNN model structure was experimented with, achieving 92.09% training accuracy and 68.30% validation accuracy after 20 epochs. Further fine-tuning with a different experimental setup of CNN, involving optimized batch processing and feature extraction, yielded a test accuracy of 91.80%. To further improve performance, a Siamese network was implemented and trained using the Nadam optimizer and learning rate scheduling, achieving a test accuracy of 86.53%. Despite these promising results, concerns about overfitting and the need for more extensive data augmentation remain. This study provides a foundation for Urdu emotion recognition, but more work is needed to achieve comprehensive emotion understanding in low-resource languages.