Fusing audio and visual data on multi-modal approaches has become a powerful solution to improve the accuracy and robustness of biometric recognition systems, particularly in real-word scenarios where single-modality systems face limitations. Additionally, cross-modal matching has emerged as a valuable approach in scenarios where direct comparisons between voice and face are impractical. While deep learning has significantly advanced biometric recognition, the computational complexity of state-of-the-art models hinders their deployment in resource-constrained environments. Recently, efficient architectures offer a promising solution, but their effectiveness in multi-modal and cross-modal scenarios has not been fully exploited. In this paper, we assess the performance lightweight neural networks across a comprehensive experimental evaluation on the standardized VoxCeleb dataset. The obtained results demonstrate that lightweight models can achieve competitive verification rates while drastically reducing computational overhead, making them viable for deploying in edge devices and real-word applications.