Tone and intonation modeling in African tone languages

This dissertation puts forth a unified model of tone and intonation representation, called the SubTonally Reductive and Additive Model (STReAM), which uses autosegmental representations. STReAM solves several outstanding issues in the representation of tone, including but not limited to (a) the Natural Class problem, i.e. the idea that there is a lack of motivation for natural classes along subtonal features, (b) the undergeneration/overgeneration problem, whereby existing subtonal feature systems predict either less or more than the maximum of five tone levels empirically attested in natural languages, and (c) the failure of previous models to capture the additive and subtractive effects of tonal processes (Hyman 2010; Clements 2010). While convincing empirical arguments for (a) have been recently offered (Meyase 2021; Akumbu 2019; McPherson 2017), all existing tonal feature models (Yip-Pulleyblanck's [+/-upper, +/-raised] and its variations) fail to account for languages with four tones in which three out of the four tones form a natural class (e.g. Fe'Fe' Bamileke and Boko; Odden (1995)). STReAM not only precisely captures the processes in these languages, but also specifically predicts the direction of pitch modification in the said processes. Issue (b) is a non-issue since STReAM only generates the maximum of 5 levels of contrast. As for issue (c), it is trivially addressed because STReAM's features are additive when they have matching signs and cancel out when they have opposite signs. STReAM is therefore a more descriptively adequate model of tonal representation than any such model previously proposed. STReAM uses two features +1 and -1, and two types of geometrical structures, a simplex one which accommodates a single feature (i.e. +1 or -1) and a complex one with two features (i.e. +1|+1, -1|-1, or +1|-1), yielding the total of five levels of contrast alluded to above. The polarity signs of the features indicate the up (+) and down (-) movements of pitch relative to an abstract reference line or register (r); the integer 1 is a unit of step, i.e. how much the pitch moved up or down. Following Jardine (2016) (also see Goldsmith (1976); Coleman & Local (1991), and others), the proposed representation is formalized as graphs. The formal novel contribution of this dissertation is in recasting all tonal processes as ones abiding by the Connectedness constraint, which is the requirement that all surface tonal representations involved in a given process be connected to each other. Since satisfying Connectedness leads to nodes from two different tones to be connected, which is also spreading, every tonal process thus becomes a spreading process. It then falls out that the difference between, say, downstep and `tone doubling' is a matter of what nodes are connected. Tone and intonation warrant a shared representation since they are both contrastive usages of pitch, only at different levels of the grammar. For this purpose, STReAM is extended to represent and model intonation as well. Register (r) is reintroduced into the representation in order to capture global, domain-wide intonational effects (Chapter 3). Statistical analysis of novel intonational data from Ede Chaabe is used to provide empirical support for register as an intonational feature bound to prosodic domains (Chapter 3 & 4). Along with the tonal representation component step (s), r forms two of the five parameters of STReAM's intonational model (presented as an algorithm); the other three being phonetic parameters for declination (d), a speaker variability parameter (alpha), and a meta parameter MCMD used to compute the others. With this handful of parameters, the STReAM algorithm is able to predict the surface intonation of different types of utterances in a variety of tone languages with very high accuracy. STReAM further allows for the ability to tell if a language with four tones can accommodate structural intonation given how crowded its tonal space is. The quantitative data from the language with four tones, Baatonum, establishes that a language with four tones can in fact have structural intonation as predicted by STReAM (Chapter 5), contrary to conjectures in the literature (Connell 2017; Downing & Rialland 2018).