**5. General Discussion**

We analyzed distributions of the grammatical, lexical, and sublexical varieties in spontaneous conversational speech produced by 40 speakers of American English [32] to assess the effects of the statistical structure of speech on the sublexical variance observed in the signal. Our results show that distributions of regularities in co-occurrence patterns, the lexical contrasts they discriminate between, and the sublexical variety observed in the articulated forms result in distributions which are consistent with previous, similar analyses of written English that satisfy many of the communicative constraints described by information theory [3].

Accordingly, these results also provide further evidence that power law distributions seen in aggregate word frequency distributions are product of mixing functionally relevant distributions that are in themselves geometric [3,8].

The distributions in the analyzed sample sugges<sup>t</sup> that, unlike the codes in artificial communication systems, human speech is a highly structured system of nested communicative distributions shaped by learning. In line with the predictions of learning theory, this suggests that speech variation at positions of high uncertainty is driven by the interaction of regular structures at multiple levels of description and that this variance serves to increases the efficiency of communication by increasing the amount of contrast in signals.

Taken together, our results indicate that the variance in the pronounced forms systematically structures the uncertainty discriminated by communicative contexts, supporting the suggestion that empirical distributions of phonetic contrasts in speech are components of a larger, highly structured communication system.

**Author Contributions:** conceptualization, M.L. and M.R.; methodology, M.R. and M.L.; verification, M.R.; formal analysis, M.L.; writing—original draft preparation, M.L. and M.R.; writing—review and editing, M.R. and M.L.; visualization, M.L.; supervision, M.R. All authors have read and agreed to the published version of the manuscript.

**Funding:** We acknowledge support by Open Access Publishing Fund of University of Tübingen. M.L. was funded by an European Research Council (ERC) advanced gran<sup>t</sup> to Harald Baayen under gran<sup>t</sup> agreemen<sup>t</sup> No. 742545.

**Conflicts of Interest:** The authors declare no conflict of interest.
