Observed effects of "distributional learning" may not relate to the number of peaks. A test of "dispersion" as a confounding factor

Distributional learning of speech sounds is learning from simply being exposed to frequency distributions of speech sounds in one's surroundings. In laboratory settings, the mechanism has been reported to be discernible already after a few minutes of exposure, in both infants and adults. These "effects of distributional training" have traditionally been attributed to the difference in the number of peaks between the experimental distribution (two peaks) and the control distribution (one or zero peaks). However, none of the earlier studies fully excluded a possibly confounding effect of the dispersion in the distributions. Additionally, some studies with a non-speech control condition did not control for a possible difference between processing speech and non-speech. The current study presents an experiment that corrects both imperfections. Spanish listeners were exposed to either a bimodal distribution encompassing the Dutch contrast /ɑ/∼/a/ or a unimodal distribution with the same dispersion. Before and after training, their accuracy of categorization of [ɑ]- and [a]-tokens was measured. A traditionally calculated p-value showed no significant difference in categorization improvement between bimodally and unimodally trained participants. Because of this null result, a Bayesian method was used to assess the odds in favor of the null hypothesis. Four different Bayes factors, each calculated on a different belief in the truth value of previously found effect sizes, indicated the absence of a difference between bimodally and unimodally trained participants. The implication is that "effects of distributional training" observed in the lab are not induced by the number of peaks in the distributions.