Production and Convergence of Multiscale Clustering in Speech

Language entails the scaling of variability across levels of measurement—small linguistic variations occur at the millisecond level, larger variations occur at the next level, and even larger variations occur over longer timescales. For acoustic onsets in speech signals, small temporal variations occur at the phonetic level, larger variations occur at the phrasal level, and even larger variations occur at the conversational level. Scaling across levels of measurement can be quantified in terms of power law distributions. In this article we review recent investigations into power law clustering of acoustic speech onsets. Studies demonstrate that the multiscale clustering in onsets reflects communicative aspects of speech in adult conversations as well as infant vocalizations. We also review evidence that multiscale clustering in the vocalizations of individuals converges during vocal interactions. We relate multiscale convergence to the notion of complexity matching, that is, the hypothesis that maximal information transfer occurs when the power laws of 2 interacting complex systems are matched. We conclude by discussing potential extensions of this work including estimating the multifractal structure of speech and testing the maximal information transfer prediction of complexity matching.