Audio-visual speech enhancement concerns a multimodal task that aims to provide a clean reconstruction of a speech given visual information and noisy audio signals. The assignment is particularly attractive for medical purposes since it might provide resources to aid deaf individuals, besides being useful for distinct contexts like social interactions in uproarious environments. This paper proposes RecognAVSE, an audio-visual speech enhancement solution developed for the AVSEC-3 challenge that combines Separable 3D CNNs and Deep Complex U-Nets to create an efficient alternative to tackle the problem. The model learns correlated features between the noisy audio signal and the visual stimuli, thus inferring meaning to the speech by amplifying relevant information and suppressing noise. Experiments over the AVSEC-3 dataset show that RecognAVSE can obtain outstanding results, outperforming the baselines in quantitative and qualitative results.