Research

Whether phonetically natural patterns are more learnable than phonetically unnatural ones remains an open question. This project addresses this debate through a case study of echo reduplication in Malay (e.g., batu-bata ‘brick’). Following Ryan (2019), head-initial languages such as English and Malay exhibit a prosodic end-weight effect, whereby prosodically heavier constituents are preferred in final position to better align with the language’s prominence system. Drawing on both a corpus study of attested Malay echo-reduplicated forms and a nonce-word judgment experiment, I found that patterns conforming to the end-weight effect (i.e., phonetically natural patterns) were indeed more learnable than unnatural ones when lexical frequency was controlled. However, the results also show that learners can acquire phonetically unnatural patterns when such patterns are sufficiently frequent in the lexicon. I capture this interaction between phonetic naturalness and lexical frequency using a probabilistic constraint-based model, e.g., Maximum Entropy Harmonic Grammar (MaxEnt), which implements asymmetric Gaussian priors over constraints according to their phonetic naturalness (Wilson 2006; White 2017).

In San Martín Peras Mixtec (SMPM; Oto-Manguean), a high tone preceded by a low tone at a noun–adjective boundary may optionally change into a rising tone (/LNoun + HAdj/ → [L + LH]). Two phonetic factors increase the likelihood of this tonal raising in the lexicon: vowel height and phonation type. Specifically, a low vowel and/or a glottalized vowel favor a rising tone more than a non-low and/or modal vowel. These factors are phonetically grounded: low vowels have intrinsically lower fundamental frequency (F0) and longer duration than non-low vowels, and glottalization lowers F0 in SMPM, both facilitating tonal raising. However, we show, in an auditory nonce-word judgment task, that native speakers of SMPM acquired the preceding-tone effect but not the vowel height or phonation effects. In other words, dependencies between tones seem to be more easily learned than dependencies between tones and vowels, suggesting the presence of an analytic bias that constrains the range of possible phonological interactions (Moreton 2008). 

In languages where word-final voiced stops are prohibited, final devoicing is the dominant repair strategy. Yet, given a hypothetical form like /tab/, several alternative repairs are conceivable, e.g., nasalization ([tam]), deletion ([ta]), vowel epenthesis ([tabə]), or metathesis ([bat]). This is known as the Too-Many-Solutions Problem and challenges a central tenet of classic Optimality Theory (OT): that free constraint reranking should yield distinct grammars, each corresponding to a different repair. Steriade’s (2001) P-Map hypothesis (“P” for perceptibility) offers a solution, proposing that languages prefer repairs that minimize perceptual distance between input and output forms. Under this view, final devoicing is the most perceptually similar repair available. While the P-Map hypothesis has been highly influential and widely cited, its core claims have not yet been tested experimentally. My dissertation addresses this gap through a series of speech perception studies (AXB) and artificial language learning experiments. Results from 127 American English speakers generally support the P-Map hypothesis, but only under certain circumstances, namely when a marked structure arises from perceptual difficulty (e.g., word-final voiced stops). When a marked structure stems from articulatory difficulty (e.g., post-nasal voiceless stops), P-Map predictions fail to hold. This asymmetry mirrors typological patterns: avoidance of post-nasal voiceless stops involves a conspiracy of repair processes (Pater 1999/2004), whereas avoidance of word-final voiced stops does not. These findings highlight the need to disentangle perceptual and articulatory motivations underlying phonological patterns and typological asymmetries. 

Paiwan, an Austronesian language spoken in southern Taiwan, prohibits the occurrence of two labial consonants in close proximity. When such sequences arise through -m- infixation (e.g., /b-m/), the infixal labial nasal surfaces as a lingual nasal ([b-n]). One potential explanation for the origin of such dissimilatory patterns is Ohala’s (1993) perceptual hypercorrection hypothesis, which proposes that dissimilation arises when listeners misattribute a coarticulatory feature (in this case, labiality) to a neighboring source. To test this hypothesis, we synthesized a continuum from [m] to [n] using the STRAIGHT speech synthesis tool (Kawahara et al., 2008) and embedded these stimuli in words beginning with either a labial ([m, b]) or alveolar ([n, d]) consonant. If perceptual hypercorrection is responsible for labial dissimilation, labial-initial contexts should elicit fewer [m] responses than alveolar-initial ones. However, results from a fieldwork-based perception experiment with 26 Paiwan listeners provided no support for this prediction. On the contrary, labial consonants tended to promote harmonic rather than dissimilatory responses, although such harmony was observed only when consonant identity was matched ([m] but not [b]).

In nearly all modern English dialects, the orthographic sequence <ng> has merged into a single velar nasal [ŋ]. Some scholars have proposed that this coalescence is a relatively recent innovation, emerging during the late Modern English period (Bermúdez-Otero 2006, 2011; Bermúdez-Otero & Trousdale 2012; Bailey 2021). Drawing on multiple historical data sources, including the Dictionary of Old English (DOE), the Linguistic Atlas of Early Middle English (LAEME), and the Middle English Dictionary (MED), we demonstrate that NG-coalescence has far deeper diachronic roots. Orthographic and rhyming evidence both support this claim: for instance, the final <g> in cyning ‘king’ and þing ‘thing’ is sometimes omitted in the DOE and MED respectively, and pre-1400 rhyme pairs such as Mapyne and endinge point to the same development. We argue that the origin of NG-coalescence must be traced back at least to Middle English and that the process was catalyzed by vocalic loss in inflectional paradigms. Once the inflectional vowel was lost, <ng> no longer occupied a prevocalic position, a context providing crucial perceptual cues to [g]. In post-consonantal and pre-pausal/consonantal contexts, by contrast, [g] became perceptually vulnerable and thus more prone to phonetic reduction and eventual deletion.

In Malay, nasality spreads iteratively and rightward from nasal consonants to following vowels and glides, but is blocked by supralaryngeal consonants. In reduplicated forms, Onn (1976) reported overapplication of nasal spreading (e.g., [w̃ãŋĩ-w̃ãŋĩ]): the first syllable of the reduplicant (the underlined constituent) acquires nasality even though there is no local trigger preceding it. This pattern carries significant theoretical implications because only parallelist (McCarthy & Prince 1995) but not serial/derivational theories of reduplication (Inkelas & Zoll 2005; Kiparsky 2010; McCarthy et al. 2012) can account for it. We collected acoustic data from 30 native speakers of Malay and found that nasal spreading in Malay displays substantial variation both within and across individuals. In reduplicated words such as /abaŋ-abaŋ/ ‘brothers’, all possible combinations of oral and nasal realizations were attested, including underapplication ([abaŋ-abaŋ]), normal application ([abaŋ-ãbaŋ]), unmotivated “pathological” application ([ãbaŋ-abaŋ]), and crucially, overapplication ([ãbaŋ-ãbaŋ]). Of these, overapplication emerged as the most frequent variant, corroborating Onn’s (1976) descriptive observations and supporting parallelist theories. The same study also revealed a phonetic correspondence effect, whereby vowels in the reduplicant and base tend to exhibit matching degrees of nasality. To account for these variable and gradient patterns, we developed a constraint-based model within the framework of generative phonetics, in which constraint violations are assessed scalarly rather than categorically (Flemming 2001; Lefkowitz 2017). The model yielded a good overall fit to the experimental data, demonstrating how integrating phonetic detail into a formal grammar can shed new light on complex issues in the morphology-phonology interface.

Based on morphosyntactic and semantic evidence, Soh and Nomoto (2011: 90) argue that Malay causative construction has the hierarchical structure shown on the left, where the verbal root first merges with the causativizing suffix -kan, and the resulting constituent subsequently combines with the verbal prefix məN-. However, I present evidence from phonology and prosody in favor of the alternative structure shown on the right, in which prefixation precedes suffixation. First, several phonological processes, such as nasal coalescence and nasal place assimilation, apply at the prefix-stem boundary but not at the stem-suffix boundary. Second, the syllable immediately preceding the -kan suffix consistently bears a high tone, which is absent at the prefix-stem juncture. This high tone has been described as marking the right edge of an Accentual Phrase in Malay (Hamzah & German 2014). Together, these observations suggest that the stem-suffix boundary is prosodically larger than the prefix-suffix boundary. Assuming some degree of  isomorphism between morphosyntax and phonology/prosody, these findings align only with the structural representation shown on the right, thereby giving rise to a bracketing paradox. 

In a six-month consultant work with a native speaker, I show that the triggering of continuancy assimilation in Bilin (Central Cushitic) is sensitive to morpheme identity. Both singular masculine nouns and plural nouns ending in a consonant take the suffix -si when marked for the accusative case. However, only the singular masculine accusative marker (-siSgM) triggers assimilation of a stem-final alveolar stop to [s] (/t + -siSgM/ → [ssi]), whereas the plural accusative marker (-siPl) does not (/t + -siPl/ → [tsi]), even though the two suffixes are segmentally identical. Interestingly, the fake geminate that results from assimilation with the singular masculine marker is durationally shorter than a sequence of two underlying [s] segments (/s + -siSgM/Pl/ → [ssi]). This suggests that underlying phonological information is still preserved within the phonetic module, resulting in incomplete neutralization. 

More info coming soon!

More info coming soon!