The Prosodic Word (Pwd) is a foundational notion in phonological theories, being relevant for the statement of many phonological generalizations. In spite of this importance, basic open questions remain about prosodic words. How many distinct prosodic categories, even at the ‘word’ level, are available to languages? Are they innately determined or emergent? Can their structure in one language vs. another be predicted? In this paper I consider a research program that attempts to address such questions by viewing prosodic words as emergent over time from the interaction of phonetics, phonologization, and syntactic structure. An important bridging idea in this understanding of prosodic words is that of domain generalization.

Phonetic dispersion has been proposed as an explanation for a number of sound-change phenomena, including vowel chain shifts, compensatory sound change, and universal trends within phoneme inventories. These explanations usually take the form of speaker-based accounts (principally H&H theory; see Lindblom, 1986). But there is a lack of empirical evidence for speaker-based approaches (McGuire and Padgett, 2011), which have also been criticized as being teleological (e.g. Blevins, 2004). This thesis explores an alternative, listener-based, perceptual account of dispersion: the Filtering Listener (Labov, 1994: 587; Wedel, 2006; McGuire and Padgett, 2011). Based in exemplar theory (Nosofsky, 1986; Pierrehumbert, 2001; etc.), the Filtering Listener hypothesis argues that when listeners are confronted with phonetically ambiguous percepts, they may not store them to phonetic memory. In turn, these unstored percepts do not update the phonemic categories of the listener, and are thus not reflected in that listener's future productions. This thesis explores the Filtering Listener as a mechanism for contrast maintenance and source of sound change, and tests its predictions experimentally in a perceptual recognition task. Participants heard phonetically ambiguous target words and phonetically unambiguous control words in noise and were asked to identify them, following Goldinger (1996). This task was repeated in 4 identical experimental blocks. Results showed that participants improved their accuracy of recognition much more over the course of the experiment for the unambiguous condition, suggesting that their storage of ambiguous percepts was degraded in memory. This provides promising preliminary evidence for the Filtering Listener; further replication, with more tightly controlled pilot studies, is required.

This dissertation begins with the observation of a typological asymmetry within phonological patterns related to phonetic naturalness. Patterns that are rooted within existing tendencies of perception and/or production -- in other words, patterns that are phonetically "natural" -- are common in phonological typology and seen in a variety of unrelated languages. Phonetically "unnatural" patterns, which are not based in these perceptual and/or production tendencies or which actively work against them, are either typologically rare or unattested. This observation has been noted in numerous works in phonology, and the debate over the nature of this asymmetry has spanned the past few decades. This dissertation makes a contribution to this debate.

There are two major arguments as to what underlies this observed typological asymmetry. One argument suggests that the bias against unnatural patterns is due to constraints on pattern learning. Learners have an innate knowledge of phonetic factors, and this knowledge predisposes them towards learning natural patterns and away from learning unnatural patterns. I refer to this argument as the universal learning bias account. The other argument suggests that the bias against unnatural patterns is rooted in perception and production. Since unnatural patterns work against phonetic tendencies, they are more difficult to perceive and/or produce than their natural counterparts and are more likely to be filtered out of typology through the forces of sound change. Under the channel bias account, the asymmetry can be explained by systematic errors in transmission that veer in the direction of natural patterns.

This dissertation explores predictions of the channel bias account using experimental methodology. A common practice within laboratory phonology is to test for naturalness-based asymmetries using artificial grammar experiments. Participants are assigned to learn one of two artificial patterns -- a natural pattern or its unnatural equivalent -- and then are tested on how well they learned their pattern. The results of these experiments, however, are often quite mixed: some experiments find that the natural pattern was learned more successfully, while others find no performance differences between the two patterns. This has led researchers to question how robust these naturalness effects truly are, and whether or not they can reliably be recreated in the lab.

I argue that the channel bias account holds a solution to this issue: namely, that proper presentation of experimental stimuli matters. Simply presenting stimuli to participants in slow, hyperarticulated speech is not enough. In order to reliably see an effect of naturalness, the phonetic differences between the patterns must be encoded into the design and presentation of the stimuli. This dissertation focuses on pattern pairs where the asymmetry is based in perception, hypothesizing that the natural pattern will be learned more successfully than the unnatural pattern if and only if the stimuli of the natural pattern are also perceived more successfully. In this way, the cross-experimental inconsistencies are easily predicted by the channel bias account, while the universal learning bias account fails to provide an explanation.

This dissertation uses experimental methods to argue for the channel bias account in three different phonological domains. In the domain of weight-sensitive stress, an artificial pattern that was both natural and formally simple was learned more successfully than two unattested patterns: one which was unnatural but formally simple, and one which was complex but natural. Perceptual difficulty could explain the poor performance within the unnatural pattern, but not within the complex pattern. Following this, I propose a 2x2 factorial template for testing for naturalness in the lab, wherein participants are assigned to learn one of two patterns (natural or unnatural) in one of two voice qualities (careful or casual). It is predicted that the group learning the unnatural pattern in casual speech will struggle to both perceive and learn their pattern; in careful speech, the unnatural pattern should be both perceived and learned with relative ease. Two experiments in the domains of final devoicing and coda sonority demonstrate evidence for the channel bias account using this experimental design.

There has been a recurrent debate in generative phonology concerning the inclusion of hierarchical prosodic structure in phonological representations. On one side, there are those who argue that prosodic structure plays an indispensable role in the conditioning of phonological phenomena, especially stress, intonation, and segmental phonotactics. On the other side of the divide are researchers who suggest that all such phenomena yield to empirically adequate non-structural analyses, which are independently favored by criteria of theoretical parsimony.

This dissertation focuses on one aspect of the larger debate over prosodic organization: the existence of the metrical foot. In standard conceptions of phonological structure the foot is a prosodic constituent, falling between the syllable and the word, which mediates the assignment of word-level stress. The foot obviously has no role to play in non-structural theories of prosody. Such frameworks assume that stress assignment is not dependent on prosodic constituents, but is instead directly computed over segments or syllables on a metrical grid.

The evidence brought to bear on the choice of prosodic theory has, by-and-large, been drawn from the typology of attested stress patterns. Both structural and non-structural theories of stress appear capable of modelling roughly the same range of stress systems. Consequently, many of the arguments for or against foot structure have centered on the ability of each theory to express the typology of stress assignment in a compact, elegant, and predictive way. This dissertation expands the terms of the debate by examining foot-conditioned phonotactics, and to a lesser extent foot-conditioned morphology, as a window on the nature of prosodic constituency in natural language.

A major conclusion of this dissertation is that hierarchical prosodic structure must be admitted as part of phonological representations in order to capture the full range of prosodically-conditioned segmental phonotactics found in natural language. Three case studies form the heart of this claim. Specifically, I show that Huariapano, Uspanteko, and Irish all manifest foot-dependent phonotactics that cannot be insightfully analyzed without recourse to abstract metrical structure. Importantly, these arguments go beyond claims about relative theoretical elegance: non-structural analyses simply fail to account for the relevant phonological phenomena in an explanatory way.

In arguing for a foot-based theory of stress assignment, I also make the case for a fairly traditional conception of the metrical foot. First, I contend that stress is always assigned on the basis of foot structure, and only to foot heads, though feet may be unstressed or `covert' as well. Second, I argue that feet are always maximally binary, even in languages where stress assignment itself is ambiguous between binary and unbounded footing. Third, I propose that any given language makes use of at most one system or `tier' of metrical organization (the uniformity of footing hypothesis). In particular, I demonstrate that the rhythmic phonology of Huariapano can and should be modeled within a single, unified system of foot structure, despite previous claims to the contrary. Along the way, I identify a novel source of phonological prominence effects: the augmentation of foot-initial syllables.

The dissertation closes with an artificial grammar study exploring how language learners acquire foot-conditioned segmental phonotactics. Given the `hidden' character of abstract prosodic structure, foot-conditioned phonotactics pose an interesting learning problem. This is especially true given the recent rise of the view that all synchronic phonological knowledge is the result of inductive learning over phonetic regularities in the speech stream. The results of the artificial grammar study suggest that speakers of both English and Japanese were inclined to learn a stress-conditioned vowel phonotactic in terms of foot structure rather than stress per se. The experiment thus supports the claim that language learning is sensitive to a foot-based parsing bias that encourages the use of the foot as a general explanatory device in acquisition. These findings provide a possible explanation for why binary foot structure is found in languages lacking rhythmic, foot-based stress, and suggest that the foot may be a prosodic universal.

Vowel insertion commonly occurs cross-linguistically to break up consonant clusters, particularly in loanwords. The most familiar form of this insertion is the categorical, phonological process of vowel epenthesis, which repairs syllable structure violations. However, gradient, phonetic vowel insertion also occurs in many languages. This phenomenon, known as vowel intrusion, occurs when speakers employ a gestural timing that results in an open transition between the consonants in the cluster. The resulting intrusive vowel is not a phonological segment, and therefore lacks gestural and durational targets.

This dissertation presents a case study of vowel intrusion in Turkish onset clusters, which occur in European loanwords. The non-lexical vowels in these clusters have previously been described as epenthetic vowels whose quality is categorically determined by the following lexical vowel, in a process of regressive vowel harmony. I present new evidence from experimental, corpus, and elicitation data, that the non-lexical vowel in Turkish onset clusters lacks a gestural or durational target, and does not form a syllable nucleus—distinguishing characteristics of intrusive vowels. Gestural and acoustic data comes from an ultrasound production study, which shows that the non-lexical vowel is gradiently present; does not achieve the same gestural and durational targets as lexical vowels; and is more affected by coarticulation.

A corpus study extends the investigation to other types of consonant clusters, and finds that the effect of the consonantal environment on the non-lexical vowel can be seen even in broad transcriptions. I also probe the syllabic and metrical status of using two studies of text-setting of /CC/ words. These studies reveal that is more variable and less likely to receive a beat than lexical vowels.

This evidence supports an analysis of Turkish /CC/ words as beginning with onset clusters. I present a coupled oscillator representation of vowel intrusion and onset clusters, using Optimality Theory and allowing the phonology to refers to both segments and gestures. I also argue that Turkish has historically been reanalyzed as an underlying vowel, but that language-internal variation, as well as increasingly prevalent knowledge of source languages, today maintain its status as an intrusive vowel.

Domain generalization is an account of certain word-final phonological phenomena, such as devoicing, in which they originate as phrase-final patterns that become phonologized and generalized by learners from the phrase level to the word level. Myers & Padgett (2014) tested this theory empirically through two artificial language learning experiments. They show that participants can learn an utterance-final obstruent devoicing pattern, given relatively short exposure, and generalize it to a word-level final devoicing rule.

Building upon Myers & Padgett (2014), this thesis explores whether the same type of domain generalization can account for syllable-final phenomena, or whether there is something special about the word that makes it the destination of generalization. Three artificial language learning experiments are presented, the third of which shows generalization from the word/utterance level to the syllable. Though syllables may differ from words in important ways, evidence from this study suggests that the syllable is accessible to the grammar for learning, generalization, and derivation of phonological rules.