The Accentual Phrases and Stops in Seoul Korean

In this on-going project, I am exploring the interdependency between the prosodic context and the realizations of the stops in Seoul Korean. In Seoul Korean, there are three kinds of stops: Lenis, Fortis and Aspirated.

These stops are realized differently depending on the position in the prosodic constituents. For example, Lenis is realized as 'strong' in the Accentual Phrase (AP) initial position and it is realized as 'voiced and lenited' when it is in the AP medial position.

  • Separability of tonal and segmental cues that signal AP boundaries

  • In an initial exploratory investigation of an unlabeled spontaneous speech data set, I found that the realization of lenis serves as a more categorical cue that separates AP-juncture from AP non-juncture, and I further found that the F0 cue is still informative as it distinguishes AP-level junctures from a higher level constituent juncture (ip/IP).

    You can find out more in my ICPhS paper [pdf] and the poster [pdf] I will present at HISPhonCog and ICPhS.

  • Perception of AP boundaries: tone and segmental cues of prosodic boundary: question disambiguation by prosody

  • In work in progress with Alessa Farinella, we are investigating what listeners pay attention to when the segmental cue and the tonal cue are in conflict: the segmental cue suggesting an AP boundary and the tonal cue suggesting no boundary, and vice versa. For a test case, we are exploring the ambiguity between a WH question and a Yes-or-No question in Korean. In Korean, the question "where did you go?" and the question "did you go somewhere?" can be string identical and disambiguated by prosody: in a WH question, the question word and the verb are in one AP, and in a YNQ question, they are in separate APs.

  • Modeling the interdependency between prosodic constituents and segmental categories

  • There is an interesting interdependence between the segmental realization and the context in which they are realized in: to infer the category of a given stop token, the listener must take into consideration where the stop token is in the prosodic constituent. At the same time, to infer where a given stop token is in the prosodic constituent, the listener might pay attention to the phonetic details of the stop token (i.e. what the segment is). This interdependency might work as an advantage for a learner learning the contrast between the segments. In this side of the project, I am working on modeling the perception and learning of this interdependency (c.f. Prosodic Analyzer account (Cho et al 2007) and Bayesian Prosody Recognizer account (McQueen & Dilley 2021))

Stress patterns: representability and learnability

In the academic year 2022-2023, I worked as a Research Assistant for Joe Pater and Gaja Jarosz. I worked on the following projects as the RA:

  • Learning Stress with Feet and Grids

  • In this study (co-authored with Cerys Hughes, Alessa Farinella and Joe pater), we compared the learning of the attested Quantity Insensitive stress languages using the Feet representation (ref) and the Grid representation (Gordon 2002). It was expected that the learner would "struggle" more to learn the languages with the Feet representation because it introduces a Hidden Structure problem(e.g., L L1 L can be (L L1) L or L (L1 L) and the learner needs to figure out the footing).

    However, with a learner (a batch gradient descent MaxEnt model (ref)) that can handle this hidden structure problem, it was found that the feet representation does not hinder the learning, in terms of the number of epochs (the number of times the learner needs to update the weights of constraints to get to the target) the learner takes to learn the language. We also found that the details of the constraint definitions are much more influential for the rate of learning.

    You can find out more in our AMP proceedings paper. [pdf] [poster]

  • SoftStress: Software for solving and learning stress patterns

  • I implemented the Linear Programming solver (Potts et al 2010) in a Google colab notebook, that can compute the weights of constraints (Harmonic Grammar), once you specify the stress pattern and the set of constraints you want to use to represent the pattern.

    It uses the same Linear Programming approach used in OT-Help, but specified for solving stress patterns.

    The same notebook also contains codes to do a gradient-descent MaxEnt learning that can handle both Grid-based representaions and Foot-based representations. This learner is the same learner used in our AMP poster/paper.

    You can try the code in this notebook: [Google Colab]

Phonetic implementation of high tone spans in Luganda

  • FPCA

  • Together with Cerys Hughes, Alessa Farinella and Kristine Yu, we investigated the phonetic implementation of high tone spans in Luganda.

    Phonologically, these High tone spans have the same surface representation: a sequence of H tones over multiple syllables; but they have different underlying representation: i.e., they come about via different phonological processes.

    Using Functional Principal Components Analysis (FPCA), we found differences in the phonetic implementation of these High tone spans, which suggests an incomplete neutralization of surface identical high tone spans.

    You can find out more in this TAI presentation (2021) and in this LabPhon poster (2022).

  • GAMMs

  • We tried a different analytic tool: Generalized Additive Mixed Models (GAMMs) to investigate the same data. We couldn't replicate the previous results we found with FPCA but we were still able to find differences between different kinds of High tone spans. We showed that the need for caution in interpreting implications for phonological representations based on small phonetic implementation differences.

    You can find out more in our ICPhS presentation and paper. [pdf]

Using the FPCA and the GAMMs, we analyzed the tone spans as trajectories of F0 (shape of lines), as opposed to targets of F0 ("points", "excursions", "delays", etc.) and our trajectory-shape analyses enrich our understanding of the very limited data available on like-tone spans arising from different spreading processes.

Laryngeal Realism in Hakha Chin

I worked on a Tibeto-Burman language called Hakha Chin, for my undergraduate honors thesis (supervised by Dr. Kelly Berkson). I investigated the speech-rate effects of the production of stop consonants in this language, using the Laryngeal Realism framework. I presented this at the UC Berkeley undergraduate linguistic symposium and I was awarded Provost’s Awards for Undergraduate Research and Creative Activity in the Social and Applied Sciences category for this project.

You can find out more in this paper. [link]

Modeling the U-shape learning curve of learning English plural morphology

Coming soon