Why do languages have the sounds they do? This project aims to explain the typology of over 3000 stop inventories with a probabilistic OT model, showing that a Maximum Entropy Harmonic Grammar (Hayes and Wilson, 2008) accurately predicts the contents of attested inventories. I show that the interaction of phonetically-based constraints in a probabilistic grammar can explain much of the structure found in the world’s stop inventories. Moreover, structural properties such as Featural Economy (Clements, 2003) are emergent under this approach, despite constraints modeling only functional pressures (Lindblom, 1986; 1990; Flemming, 1995; et seq.).

• A probabilistic model of stop inventory typology – Talk at Issues in Phonological Typology, 2025. Questions and comments welcome!

• A probabilistic typology of Proto-Indo-European stops – Poster at the The Seventh Edinburgh Symposium on Historical Phonology, 2025.
  The traditional reconstruction of Proto-Indo-European (PIE) has a three-way stop contrast consisting of voiceless (T), voiced (D) and voiced-aspirated (Dh) stops. This stop system, T-D-Dh, which excludes the voiceless aspirates (Th) is typologically is exceedingly rare, if not entirely unattested. I present a direct test, based on typological evidence, of the plausibility of different PIE inventories in MaxEnt, a probabilistic implementation of Optimality Theory (OT). I show that the most likely candidate for the synchronic PIE inventory is the four-way system, T-Th-D-Dh.

The complex four-way voicing and aspiration contrast of Indic languages like Bengali is poorly understood, both in terms of their phonetic and phonological properties. Voice Onset Time (VOT) which is typically used to characterize voicing contrasts is not sufficient to distinguish the four laryngeal categories of languages like Bengali. Moreover, it is unclear how this contrast is represented in terms of phonological features. In this series of projects, I characterize the acoustic properties of connected Bengali speech in two different registers – Infant Directed Speech (IDS) and Adult Directed Speech (ADS). I consider what these acoustic properties suggest about how this complex contrast is represented phonologically, and the phonetic dimensions along which voicing contrasts are organized in the world’s languages.

• Reconceptualizing VOT: Further Contributions to Marking 50 Years of Research on Voice Onset Time, Journal of Phonetics
  I argue that VOT should be reconceptualized as a two-dimensional plane rather than a one-dimensional continuum. This simple reformulation of VOT under which negative and positive VOT make up the complete VOT space yields a more complete description of the voicing contrasts that exist in the world’s languages.

• Acoustic Properties of the Four-Way Laryngeal Contrast in Bengali Infant Directed Speech, ICPhS 2023 proceedings paper
  This paper analyzes the acoustic properties of the four-way laryngeal contrast in Infant Directed Speech (IDS). I show that this contrast is not globally hyper-articulated in Bengali IDS, as suggested by some studies on other languages, nor is Bengali IDS breathier than ADS.

• The four-way laryngeal contrast in Bengali IDS and beyond, ICPhS 2023 poster
  This poster describes the acoustic properties of the four-way laryngeal contrast and shows that splitting VOT into its positive and negative components adequately captures this contrast. This conception of VOT also better captures the typology of languages that contrast three laryngeal categories.

• Laryngeal realist representations in Bengali, MFM 2023 poster
  This poster argues that the four-way contrast is best represented by the phonological features [voice] and [spread] in line with theories of Laryngeal Realism, based on acoustic data from Bengali connected speech.

Indic languages are reported to have a loanword adaptation pattern whereby English aspirated stops are adapted as unaspirated, even though Indic languages have phonological aspirates. This pattern is also observed in Indian varieties of English. This project describes and contrasts the acoustic properties of stops in Indian English and Southern British English, and shows that listeners of English and Hindi perceive the same stops differently — English listeners perceive aspiration in stops at shorter voice onset times (VOT), and their perception of aspiration is affected by consonant-induced f0 (Cf0). Hindi listeners, on the other hand, only perceive aspiration at longer VOT and their perception is unaffected by Cf0. Thus, this unexpected adaptation pattern can be explained by language-specific perception, suggesting phonetics may override phonology in loanword adaptation.

• A perceptual explanation for Hindi loanword adaptation, under review (with Max Meszaros)
• A perceptual explanation for the adaptation of aspirated stops in Indic languages, poster at the Meeting of the Acoustical Society of America, 2025 (with Max Meszaros)

• An acoustic study of voiceless stops in Indian English, paper in the Journal of South Asian Linguistics, 2022 (with Peter Staroverov)

• The Curious Absence of Aspiration in Indian English: The Role of Phonetics in Adaptation, WCCFL 37 proceedings paper, 2021

Infant Directed Speech (IDS) has been shown to be hyper-articulated and is thought to facilitate phonetic acquisition. This is particularly true for vowels and studies on consonant hyper-articulation have returned mixed results. This series of projects evaluate whether consonants and vowels IDS are, in fact, hyper-articulated and how phonetic categories may be learned from naturalistic data.

• Bootstrapping prosodic primitives from Infant Directed Speech, in prep (with Ann Aly, Stephanie Jacobson and Megha Sundara)
  We show that language-specific prosodic categories can be learned solely from intensity and frequency information in the target language. We show that a clustering algorithm can find three clusters in Spanish and Bengali, two languages that differ in their intonational phonologies, and that, crucially, these learned clusters map onto prosodic groups that approximate the target intonational phonologies. This suggests that a learner that pays attention to the same set of acoustic cues will discover typologically appropriate functional categories that signal the prosodic structure of their target language.
• Evaluating the learnability of vowel categories from Infant-Directed Speech, Poster at the Meeting of the Acoustical Society of America, 2022 (with Katya Khlystova, Connor Mayer, Ann Aly, Ji Young Kim and Megha Sundara)
  We test the facilitative effect of IDS by evaluating the acoustic separability of English and Spanish vowels and independently testing the learnability of vowel categories using a previously implemented Gaussian learner (Feldman et al., 2013). We find some evidence for greater separation in in both English and Spanish IDS. However, the Bayesian distributional learner is unable to learn appropriate vowel categories consistently. Overall, we find no evidence for the facilitative effect despite greater vowel separability. We also find that the learner, which was previously validated on lab-speech, is unable to handle the variation found in naturalistic speech.

• The Four-way Voicing Distinction in Bengali Infant Directed Speech, Poster at the Meeting of the Acoustical Society of America, 2022 (with Megha Sundara)
  We show that stops and vowels in Bengali IDS are not hyper-articulated, nor are they more acoustically separable. Some stops are, in fact, hypo-articulated, but not as a result of enhanced vowel contrasts. Apparent hyper-articulation of stops is an effect of the slower speech rate of IDS. This suggests that IDS may not always facilitate phonetic learning.

In this ongoing collaboration with Ben Eischens, we are analyzing the acoustic properties of non-modal phonation in San Martín Peras Mixtec. Many Otomanguean languages contrast modal and laryngealized phonation, which is realized as a modal vowel followed by laryngealization. Previous work has shown that the production of laryngealization in other Mixtec varieties varies greatly both within and across speakers (Gerfen and Baker, 2005; Cortés et al., 2023). Results from our acoustic study show that laryngealized and breathy vowels in San Martín Peras Mixtec are realized remarkably consistently within and across speakers via durational cues.

• The production of phonation type in San Martín Peras Mixtec, under revision (with Ben Eischens)
  In this paper, we describe the acoustics of phonation type in SMP Mixtec, which contrasts modal, glottalized, and breathy phonation. We find that both glottalized and breathy vowels are realized as voiceless for the majority of their duration, even in intervocalic position where such vowels are typically voiced. Our results add to the growing body of evidence that the presence or absence of voicing in glottal sounds is not a reliable indicator of their phonological representation.

In this collaborative project led by Gaby Hermon and the late Peter Cole, we documented Adonara (iso: adr), an Austronesian language spoken in eastern Indonesia. This collaboration included native speaker undergraduate students and Indonesian linguists. Audio and video recordings are available on PARADISEC with transcription, glosses and English and Indonesian translations.

In this series of projects, I consider how music and song-writing can inform our understanding of phonology.

• Hindi-Urdu pop text-setting
  In this ongoing collaborative project with Muhammad Rehan and Dhanya Charan, we are investigating Hindi-Urdu musical text-setting for evidence of lexical stress. Although Hindi-Urdu is described as having lexical stress, experimental studies have failed to find consistent acoustic correlates and perceptual cues that signal stress. We are analyzing recent Hindi-Urdu pop songs to evaluate the relationship between metrical and syllable structure on beat duration in text-setting. Our preliminary results show that although there is no relationship between putatively stressed syllables and their beat duration, heavier syllable onsets are set to longer beats. This suggests an intricate role of syllable structure on how words are set to music, despite no such role for stress. We also find evidence for polysyllabic shortening, a negative correlation between the length of a word and the duration of syllables within it, observed in many disparate languages to maintain speech rhythm.
• Half-rhymes in English pop song-writing
  In this project, I analyzed 200 pop songs from the 1960s to the late 1980s – 100 of which were included in Rolling Stone magazine’s 1988 list “100 Best Singles Of The Last 25 Years” and the other 100 were judged to be “The Worst 100 Singles of the Last 25 Years” by the New York Daily News in 1989. Investigating the makeup of imperfect rhymes, I found that song writers tend to favor perceptually similar sounds to form imperfect rhymes, as predicted by Steriade’s P-map, regardless of whether the songs are considered good or bad by listeners. Moreover, listeners judge rhymes as being bad when they contain semantic oddities, non-sequiturs and contradictions. While perceptually dissimilar alternations by themselves do not influence listeners’ judgement of rhyme-goodness, the additive effects of perceptual dissimilarity with semantic oddity and syntactic ill-formedness do influence ratings.