The Berkeley Phonetics, Phonology and Psycholinguistics Forum ("Phorum") is a weekly talk and discussion series featuring presentations on all aspects of phonetics, phonology, and psycholinguistics. We meet on Fridays from 4(:10)-5pm (unless specified otherwise below), in Dwinelle 1229 (Zoom link shared upon request). Phorum is organized by Kai Schenck. My email is "kai_schenck" @berkeley.edu.
Schedules from previous semesters can be found here.
Spring 2026
February 13
Kai Schenck (UC Berkeley): Unsupervised learning at the limits of phonology: unbounded circumambient patterns
Formal language theory (FLT) is able to quantify the computational complexity of phonological processes by treating them as string-based input-output mappings that are accepted by some formal grammar. Although non-optional phonological mappings have long been known to be a subset of the regular input-output mappings (Heinz 2011a), the attested phonological processes which require the most complexity to model are known as unbounded circumambient (UCA) – requiring information from both sides of a undergoer that may, in principle, be unboundedly far away (Jardine 2016, McCollum et al. 2020, Meinhardt et al. 2024). However, the unattested Sour Grapes harmony pattern (Wilson 2003a, 2006) also falls into this class (Lamont 2019, Meinhardt et al. 2021, Yolyan 2025). Input-output mappings based in FLT thus are unable to cleanly separate attested from unattested processes – various explanations have been proposed to account for this discrepancy such as phonotactic complexity (Lamont 2019) or diachronic stability (Prickett 2025). While several studies investigate the learning of Sour Grapes compared to a simpler attested process (Finley 2008, Lin & Myers 2010, Lai 2022, Prickett 2022, 2025), no studies compare either to attested UCA processes.
I investigate whether Sour Grapes harmony is more difficult to learn compared to attested UCA processes by training Generative Adversarial Networks (GANs) on artificial language input. GANs have been used extensively in work on the learning of phonological alternations (Beguš 2020, 2021b, Barman et al. 2024) and
are general learners with no language-specific substantive bias (Beguš, Zhou, et al. 2023), making them ideal to test whether Sour Grapes harmony is inherently more difficult to learn despite its similar classification to attested processes in FLT. I train three GANs, each on an artificial language with the same input forms that had a different output harmony system: one language with a simpler bidirectional harmony system, one that resembles attested UCA processes, and one that implements Sour Grapes. I find that, although there are small indicators that the bidirectional system was learned more easily, there is no statistically significant difference in learning of all three systems when only tokens with active spreaders are examined. This is in line with some previous some results for humans (Lin & Myers 2010, Prickett 2022), suggesting both that the formal complexity of Sour Grapes is not responsible for its unattestedness and that a general learner such as a GAN learns both simpler harmony processes and UCA processes to approximately the same level regardless of their attestedness.
February 20
Noah Macey (UC Berkeley): What makes an accent 'foreign'? Preliminary evidence from self-supervised speech models.
This preliminary work investigates whether and how L1 and L2 accents differ systematically. As a motivating example, imagine that a linguist meets several people with accents she has never heard before. Without asking for any biographical information, would she be able to tell which speakers learned English as a first language?
There are three possibilities for how the linguist would perform, corresponding to three hypotheses about L1 and L2 accents. First, if our linguist makes the L1 vs. L2 determination with near 100% accuracy, then we can say that L1 and L2 accents are categorically different. Second, if our linguist is no better than chance at distinguishing L1 vs. L2, then we can say that there is no acoustic difference between L1 and L2 accents (though obviously there are social differences). Equivalently, we could say that every L2 accent is a potential L1 accent---there is no way to distinguish between them without resorting to biographical information. Third, if our linguist makes a better-than-chance but worse-than-certain judgement about each speaker, then we can say that L1 and L2 accents cluster together in accent-space, with significant overlap, but that there is no hard-and-fast acoustic difference between L1 and L2 speech.
The research presented here contributes to this question by comparing speech-model representations of different accents. The data comprise standardized phrases from the George Mason University Speech Accent Archive (Weinberger, n.d.), representing 1,204 English speakers of various backgrounds. Following Chernyak et al. (2024), an algorithm from the HuBERT family (Hsu et al. 2021) is used to embed the accent recordings in an acoustic space. Once embedded, the recordings are interpretable as trajectories through a three-dimensional space. These trajectories are clustered to test if L1 and L2 accents group together, and a k-nearest neighbors classifier tests whether L1/L2 category is predictable. Finally, inter- and intra-accent variance is tested as well, in an attempt to replicate the findings of Chernyak et al. (2024). The clustering, classifier, and variance analyses fail to distinguish L1 and L2 accents, providing weak evidence against systematic acoustic differences between L1 and L2 speech.
February 27
Larry M. Hyman (UC Berkeley): A Gap in OCP Tonal Effects
In this presentation I have two goals. First, I want to present an unattested gap in how the Obligatory Contour Principle (OCP) potentially interacts with H(igh) Tone Spreading (HTS), adding to the inventory presented by Myers (1997). Second, I will present some of the facts of HTS in the Thɔnkɔ dialect of Limba, a Niger-Congo isolate spoken in Sierra Leone (Hyman & Kamara 2025a,b). The relation between the two is that the Limba facts, independently intriguing, seem to involve the gap that I have in mind [no spoiler here]. However, I will show that the facts are better treated without invoking the apparently unattested repair in question. I will begin by giving a brief refresher of how the OCP affects tone, especially in systems where H is a privative tone contrasting with Ø. I briefly ask whether the reason that the gap has not been attested has to do with the substance of tone or the formal relation (ranking) between the general OT constraints which apply to other aspects of phonology as well.
March 13
Lev Michael (UC Berkeley):
March 20
Marko Drobnjak: LLM-Assisted Welfare Language Accessibility: A User-Centered Experimental Framework
April 3
Jinyoung Jo (Stanford):
April 10
Zach Wellstood (UC Berkeley):
April 24
Larry Lyu (UC Santa Cruz):
May 1
Michelle Kamigaki-Baron (University of British Columbia):