We introduce the EM-BDD algorithm, a novel approach to learning Hidden Markov Models (HMMs) by leveraging Binary Decision Diagrams (BDDs). HMMs are pivotal in representing complex systems revolutionizing fields such as autonomous vehicles and language technology. However, the Baum-Welch (BW) algorithm, the state-of-the-art solution for learning HMMs, is hampered by its high computational complexity in each iteration, limiting its efficiency and scalability. The EM-BDD algorithm addresses these challenges by utilizing BDDs in the Forward-Backward procedure, building upon the BW algorithm’s methodology. This approach significantly affects the algorithm’s time and memory complexity, contingent on the BDD’s size, which in turn is highly dependent on variable ordering - an NP-complete problem. Our evaluation through a comparative analysis between the EM-BDD and BW algorithms demonstrates the former’s potential for enhanced efficiency and scalability in learning HMMs.