Signal flow graphs are a graphical model of signal transducers, which play a foundational role in control theory and engineering. In this work, we develop a learning algorithm for closed (i.e. with no inputs) signal flow graphs, which are behaviourally equivalent to weighted finite automata on a singleton alphabet. Analogously to the case of automata learning, our algorithm constructs a signal flow graph from a given set of output behaviours. We demonstrate that this procedure results in a genuine reduction of complexity: our algorithm fares better than existing learning algorithms for weighted automata restricted to the case of a singleton alphabet.