Accurate and objective assessment of embryo development is a cornerstone of successful in vitro fertilization (IVF), yet current practice remains highly dependent on manual and subjective evaluation of morphokinetic events. This work introduces a deep learning framework leveraging spatio-temporal transformers to automatically detect developmental phase transitions in human embryos from time-lapse imaging. We formulate this task as a binary sequence classification problem and benchmark a transformer-based architecture (TimeSformer) against a strong convolutional baseline (frame-stacked ResNet-18). Using an embryo-level split on a curated time-lapse dataset, TimeSformer achieved a substantial performance gain, reaching an F1-score of 90.44% compared to 71.82% for the CNN baseline. These results demonstrate the critical importance of modeling temporal dependencies and highlight transformers as a state-of-the-art solution for morphokinetic analysis. Beyond outperforming conventional CNNs, our approach lays the groundwork for robust, automated, and clinically relevant decision-support tools, advancing the integration of AI into IVF workflows.