The Berkeley Phonetics, Phonology and Psycholinguistics Forum ("Phorum") is a weekly talk and discussion series featuring presentations on all aspects of phonology, phonetics and psycholinguistics. We meet on Fridays from noon to 1 pm. We plan to have a virtual option available for those who would prefer to join virtually: please email one of the organizers for the Zoom link, or to ask to be added to the mailing list (which will include relevant links). Phorum is organized by Katie Russell and Maks Dąbkowski. Our emails are respectively "katherine.russell" and "dabkowski" @berkeley.edu.
Fall 2021 Schedule
David Gaddy (UC Berkeley): Decoding Silent Speech with Electromyography
In this talk I will discuss my research on decoding silent speech. This work uses machine learning to recognize silently mouthed words and turn them into audible speech, based on electrical signals from muscles captured on the surface of the face and neck. The talk will give an overview of silent speech and its applications, describe the machine learning models we use to decode the signals, and discuss our analysis looking into what speech features the system captures.
Jon Rawski (San José State University): Abductive Learning of Phonotactic Constraints
Inductive learning of phonotactic knowledge from data often relies on statistical heuristics to select plausible phonotactic constraints, such as in the popular Maximum Entropy learners (Hayes & Wilson 2008). Wilson & Gallagher (2018) claim that such statistical heuristics are necessary, given that feature-based constraints allow for exponentially large hypothesis spaces. I show that such statistical heuristics are unnecessary, by providing a series of non-statistical algorithms which use abduction to select the most general feature-based constraint grammars for both local and long-distance phonotactics. I compare these algorithms to MaxEnt grammars to showcase their similar behavior on synthetic and natural phonotactic data. Like any algorithms, these help us clarify general properties of phonotactic learning: 1) the space of possible constraints possesses significant structure (a partial order) that learners can easily exploit, 2) even given this structure, there are multiple pairwise incomparable grammars which are surface-true, and 3) particular constraint selection is due to the particular abductive (not inductive) principles learners possess which guide the search, regardless of whether such principles are statistically formulated or not.
AMP practice talks
Yevgeniy Melguy (UC Berkeley): Mechanisms of listener adaptation in perceptual learning for speech
Listeners often show processing difficulty when faced with a novel accent. However, research has shown that they can rapidly adapt, resulting in attenuation or disappearance of this "accent cost". The goal of this study was to gain a better understanding of the underlying mechanisms in the perceptual adaptation process. To do this, a well-established phonetic learning paradigm was utilized where listeners are exposed to an artificial accent involving an ambiguous pronunciation of a target sound (e.g., /θ/ = [θ / s]) and subsequently tested on categorizing a phonetic continuum between these two sounds. If learning is successful, listeners show a shift in their categorization boundary, such that they accept a greater proportion of sounds as instances of the trained category. Following up on earlier research (Zheng & Samuel 2020), this experiment investigated whether such phonetic learning is the result of 1) a shift of the trained phoneme category in phonetic space or 2) relaxation of categorization criteria (the expansion of the trained category in phonetic space). Results suggest that perceptual learning for speech is better explained as category shift -- listeners demonstrated specific adjustments to the accent but did not show evidence of generalizing learning to neighboring phonetic space.
Zion Mengesha (Stanford University)
Evelin Balog (Friedrich-Alexander Universität Erlangen-Nürnberg; Fulbright at UC Berkeley)
Natalie Weber (Yale University)
Chloe Willis (UC Santa Barbara)
Richard Bibbs (UC Santa Cruz)
Josefina Bittar Prieto (UC Santa Cruz)
Caitlin Smith (UC Davis)
Zachary O'Hagan (UC Berkeley): Verbal Reduplication in Caquinte