Modelling Sonority with Complementary Symbol and Signal Based Models
Total Page:16
File Type:pdf, Size:1020Kb
The double life of language systems: Modelling sonority with complementary symbol and signal based models Aviad Albert1 & Bruno Nicenboim2,3 1 Department of Linguistics—Phonetics, University of Cologne 2 Department of Cognitive Science and Artificial Intelligence, Tilburg University 3 Department of Linguistics, University of Potsdam Abstract Sonority is a fundamental notion in phonetics and phonology, central to many descriptions of the syllable and evidently useful as a predictor of phonotactics (i.e., the propensity of different consonants and vowels to com- bine via concatenation). Although widely-accepted, sonority lacks a clear basis in speech articulation or perception, and traditional sonority princi- ples exhibit gaps in empirical coverage, while they are also not designed to be compatible with general cognitive capacities. Against the backdrop of traditional sonority accounts, which have been exclusively based on dis- crete and symbolic machinery to model sonority, we propose an incorpo- ration of symbol-based and signal-based models to adequately account for sonority with two complementary models. We claim that sonority is primar- ily a perceptual phenomenon related to the strength and quality of pitch per- ception, postulating a universal drive to optimize syllables as pitch-bearing units. We suggest a measurable acoustic correlate for sonority in terms of periodic energy, and we provide a novel principle that can account for syl- labic well-formedness based on principles of competition in real-time; the Nucleus Attraction Principle (NAP). We present two perception experiments that test our two NAP-based models against four traditional sonority models and we use a Bayesian data analysis approach to test and compare the differ- ent sonority models. We show that our two NAP models retain the highest degree of complementarity while one of them is clearly superior to all the other models we tested. We interpret the results as providing strong sup- port for our proposals: (i) the designation of periodic energy as sonority’s correlate; (ii) the incorporation of continuity in phonological models, and; (iii) the dual-model strategy that separates and integrates symbol-based top- down processes and signal-based bottom-up processes. Keywords: Sonority; Periodic energy; Bayesian data analysis; Phonetics– Phonology interface Word count: X 1 Introduction The study of the sound system of human languages has been one of the longest- standing intersections of symbol-based categorical analyses on the one hand, and signal- based continuous descriptions on the other. These two different types of analysis stand at the core of the distinction many researchers make between phonetics and phonology, where the former addresses continuous and measurable aspects of the speech signal (e.g., articulation, perception and acoustics), while the latter addresses categorical aspects of the speech signal using discrete and symbolic units (e.g., consonants, vowels, and syllables). The following study proposes an incorporation of symbol-based and signal-based models to adequately account for the notion of sonority, which has played a pivotal role in the vast linguistic literature on syllables, most often invoked as a powerful tool to pre- dict phonotactics, i.e., the propensity of different consonants and vowels to combine via concatenation. We claim that sonority is primarily a perceptual phenomenon related to the strength and quality of pitch perception, postulating a universal drive to optimize syllables as pitch- bearing units (see Section 2). We suggest a measurable acoustic correlate for sonority in terms of periodic energy, and we provide a novel principle that can account for syllabic well- formedness based on general principles of competition in real-time; the Nucleus Attraction Principle (NAP). We implement NAP with two complementary models (see Section 3): (i) a bottom-up model that accepts continuous information and links it to predicted categorical responses, and; (ii) a top-down model that is based on discrete segments—consonants and vowels— and their potential distribution within syllables. In suggesting two different models for a single phenomenon we follow the pioneering works of Howard Pattee (see Pattee & Rączaszek-Leonardi, 2012) rather than the more commonly assumed Occam’s razor, which requires theories to be as simple as possible, thus creating a strong preference for a single model over two different models in explaining a certain phenomenon. In contrast, Pattee views the link between continuous dynamic modes on the one hand, and discrete symbolic modes on the other hand, as a crucial characteristic of all language systems, yet he argues further that no single model can adequately describe both the continuous and discrete modes of the system, as they are not reducible to a single mathematical representation. In section 4 we present two perception experiments—exploratory and confirmatory— that we designed in order to test our two NAP-based models (applying NAP with bottom-up and top-down approaches) against four traditional sonority models, considering two types of common sonority hierarchies (with and without distinctions between obstruents), and two types of common sonority principles—the Sonority Sequencing Principle (SSP) and the Author Note: In accordance with the Peer Reviewers’ Openness Initiative (opennessinitiative.org), all materials and scripts associated with this manuscript were available during the review process and remain available on OSF project “NAP model” at https://tinyurl.com/y3l4avap/. Declarations of interest: none. Correspondence concerning this article should be addressed to Aviad Albert, Department of Linguistics— Phonetics, University of Cologne, Herbert-Lewin-Straße 6, 50931 Cologne, Germany. E-mail: a.albert@uni- koeln.de MODELLING SONORITY WITH COMPLEMENTARY MODELS 2 Minimum Sonority Distance (MSD). We use a Bayesian data analysis approach to test and compare the six different sonority models. Whereas all the different models arefound to be capable of predicting the experimental results to some extent, the symbolic top- down version of NAP is shown to be superior and appears to exhibit the highest degree of complementarity with the bottom-up model of NAP. Our synergy of proposals has many advantages over traditional sonority accounts, including methodological aspects, theoretical perspectives, and, most importantly, a better empirical coverage. In Subsection 4.6.3 we interpret the results of the main confirmatory experiment and in Section 5 we address some relevant implications, where we discuss the usefulness of the top-down vs. bottom-up distinction instead of the common use of the phonology–phonetics dichotomy to express the distinction between discrete and continu- ous aspects of linguistic sound systems (Subsection 5.1). We also discuss the division of labor between sonority and other phonotactic factors, demonstrated with a holistic account of the phenomenon of /s/-stop clusters (Subsection 5.2), before we present our concluisons in Section 6. In the remainder of this Introduction, we briefly present the relevant background on traditional sonority hierarchies and sonority principles, emphasizing their rationale, their application, and their inherent flaws (Subsections 1.1-1.3). 1.1 Sonority hierarchies A sonority hierarchy is a single scale on which all consonant and vowel types can be ranked relative to each other. Such hierarchies can be traced back centuries, and concepts akin to sonority hierarchies can be found already in the pioneering works of early Sanskrit grammarians.1 Early versions of current sonority hierarchies often date back to Sievers (1893); Jespersen (1899); and Whitney (1865), while Ohala (1992) even goes as far back as Brosses (1765). While the phonetic basis of sonority hierarchies remains controversial, phonological sonority hierarchies have been primarily based on repeated observations that revealed sys- tematic behaviors of segmental distribution and syllabic organization within and across lan- guages. The general consensus regarding the phonological sonority hierarchy thus stems from attested cross-linguistic phonotactic behaviors of different segmental classes, such as, for instance, the relatively high frequency of stop-liquid sequences in the onset of complex syllables (e.g., /kl/ in the English word clean) and the rarity of the opposite liquid-stop sequences at such onset positions, but not at the mirror-image coda position of syllables (e.g., /lk/ in the English word milk). See examples in Zwicky (1972); Selkirk (1984); Parker (2002); Jany, Gordon, Nash, and Takara (2007), and see Ohala (1992) for related criticism regarding the circularity that results from determining sonority hierarchies according to attested behavior without another independent (phonetic) variable. 1Donegan (1978) notes that Pāṇini and the Sanskrit grammarians used the term svara to imply some kind of harmonic musical quality which applies mainly to vowels. Parker (2002, p. 58) notes further that the Sanskrit grammarians observed natural classes for speech sounds that are “grouped according to their degree of ‘opening’ (vivāra)”. MODELLING SONORITY WITH COMPLEMENTARY MODELS 3 Most common phonological sonority hierarchies group segment types into classes that primarily reflect the standard manners of articulation in traditional phonology. The distinct categories commonly used include stops, fricatives, nasals, liquids, glides, and vow- els, often with additional distinctions such as voicing