Re-evaluating Phoneme Frequencies

Abstract

Causal processes can give rise to distinctive distributions in the linguistic variables that they affect. Consequently, a secure understanding of a variable's distribution can hold a key to understanding the forces that have causally shaped it. A storied distribution in linguistics has been Zipf's law, a kind of power law. In the wake of a major debate in the sciences around power-law hypotheses and the unreliability of earlier methods of evaluating them, here we re-evaluate the distributions claimed to characterize phoneme frequencies. We infer the fit of power laws and three alternative distributions to 166 Australian languages, using a maximum likelihood framework. We find evidence supporting earlier results, but also nuancing them and increasing our understanding of them. Most notably, phonemic inventories appear to have a Zipfian-like frequency structure among their most-frequent members (though perhaps also a lognormal structure) but a geometric (or exponential) structure among the least-frequent. We compare these new insights the kinds of causal processes that affect the evolution of phonemic inventories over time, and identify a potential account for why, despite there being an important role for phonetic substance in phonemic change, we could still expect inventories with highly diverse phonetic content to share similar distributions of phoneme frequencies. We conclude with priorities for future work in this promising program of research.

The sounds of prehistoric speech

“…Regardless of the number of phonemes in a given language, the frequency of occurrence of those phonemes follows a power-law distribution whereby a few phonemes and their variants are incredibly frequent with respect to the others [20–22]. In languages tested to date, a phoneme's intralinguistic frequency, F , is predictable by its intralinguistic frequency-based rank, r , when factoring in the total number of phonemes in the language, n [20–22].…”

“…Regardless of the number of phonemes in a given language, the frequency of occurrence of those phonemes follows a power-law distribution whereby a few phonemes and their variants are incredibly frequent with respect to the others [20–22]. In languages tested to date, a phoneme's intralinguistic frequency, F , is predictable by its intralinguistic frequency-based rank, r , when factoring in the total number of phonemes in the language, n [20–22]. Several equations have been proposed to account for this predictable relationship, for instance the following relatively straightforward equation that has been shown to be accurate across many languages [20]F=log⁡false(n+1false)−log⁡rn.While debate exists regarding the best way to describe the power-law distribution of sounds in speech, and slightly more accurate formulae are derivable, there is no debate that sounds generally adhere to a power-law distribution, both in word lists and in transcriptions of discourse [20–22].…”

“…In languages tested to date, a phoneme's intralinguistic frequency, F , is predictable by its intralinguistic frequency-based rank, r , when factoring in the total number of phonemes in the language, n [20–22]. Several equations have been proposed to account for this predictable relationship, for instance the following relatively straightforward equation that has been shown to be accurate across many languages [20]F=log⁡false(n+1false)−log⁡rn.While debate exists regarding the best way to describe the power-law distribution of sounds in speech, and slightly more accurate formulae are derivable, there is no debate that sounds generally adhere to a power-law distribution, both in word lists and in transcriptions of discourse [20–22]. There is no motivation for suspecting that pre-Neolithic languages differed from contemporary languages according to this parameter either.…”

Challenges of sampling and how phylogenetic comparative methods help: With a case study of the Pama-Nyungan laminal contrast

“…Languages vary not only in what contrastive segments they have but also in how frequently they use them (Frisch, Pierrehumbert & Broe 2004, Hall 2009, Wedel, Kaplan & Jackson 2013, Macklin-Cordes & Round 2020. For example, Pitta Pitta (Blake 1990) and Burduna (Burgman 2007) are similar in that they both contrast laminal stops, nasals and laterals in word-initial position.…”

Assessing phoneme distribution for speech modeling

“…For example, Martindale et al [12] proposed that phoneme frequencies track a particular member of the power law family, using the Yule-Simon distribution as a basis [13]. A recent study discuss the idea of using power laws to recognize distributions [14], and aimed to understand the analysis of phonetic distributions as properties of variables, arguing that a variable can be defined as the set of values that characterize something. Thus, a population speech sample with a specific language could be characterized through a mathematical distribution [14]; the authors conclude the need to rigorously evaluate the fit of the distributions used to evaluate the frequencies of phoneme use in a given language.…”