Sound is a medium used by humans to carry information. The existence of this kind of medium is a pre-requisite for language. It is organized into a code, called speech, which provides a repertoire of forms that is shared in each language community. This code is necessary to support the linguistic interactions that allow humans to communicate. How then may a speech code be formed prior to the existence of linguistic interactions? Moreover, the human speech code is characterized by several properties: speech is digital and compositional (vocalizations are made of units re-used systematically in other syllables); phoneme inventories have precise regularities as well as great diversity in human languages; all the speakers of a language community categorize sounds in the same manner, but each language has its own system of categorization, possibly very different from every other. How can a speech code with these properties form? These are the questions we will approach in the paper. We will study them using the method of the artificial. We will build a society of artificial agents, and study what mechanisms may provide answers. This will not prove directly what mechanisms were used for humans, but rather give ideas about what kind of mechanism may have been used. This allows us to shape the search space of possible answers, in particular by showing what is sufficient and what is not necessary. The mechanism we present is based on a low-level model of sensorymotor interactions. We show that the integration of certain very simple and non language-specific neural devices allows a population of agents to build a speech code that has the properties mentioned above. The originality is that it pre-supposes neither a functional pressure for communication, nor the ability to have coordinated social interactions (they do not play language or imitation games). It relies on the self-organizing properties of a generic coupling between perception and production both within agents, and on the interactions between agents.
