Emergent Communication is a research field that encourages the natural emergence of communication protocols in multi-agent systems, by letting them collaborate in solving problems. This paper proposes a dynamic vocabulary scheme in which agents can autonomously expand their vocabularies when learning to communicate in a sequence of discrete symbols. This paper also shows that, by letting the vocabulary size to grow naturally, agents reach a stable communication protocol faster, requiring less training data. In addition, the emerged communication protocol is more compositional and interpretable than that of the fix-sized vocabulary scheme counterparts.