The Berkeley Phonetics & Phonology Forum (Phorum) is a weekly talk and discussion series featuring presentations on all aspects of phonology and phonetics.
What is a phonology lab? What goes on there? Can I use the facilities of the UC Berkeley Phonology Lab?
The first meeting of Phorum this semester (Sept. 13) is an introduction to the Phonology lab. The plan is to have an overview presentation, and then some demos of how people get their work done in the lab.
The talk is based on Yoneyama and Munson (2010), a first report of a study of the influence of word frequency, word familiarity, and phonological neighborhood density on spoken-word recognition in Japanese adults acquiring English. Three groups of listeners participated: lower-proficiency English learners in Tokyo, Japan, higher-proficiency English learners in Minneapolis, US, and native speakers of English in Minneapolis. Following Imai, Walley, and Flege (2005), we presented listeners with 80 words that varied orthogonally in word frequency and phonological neighborhood density, spoken by either a native speaker of English or a native speaker of Japanese producing Japanese-accented English. Robust effects of phonological neighborhood density and talker native language were found for all three groups; moreover, the high-proficiency Japanese listeners showed a smaller decrement in perception when listening to the accented talker than did listeners in the other two groups. However, a close analysis of the stimuli showed that vowels and consonants were not distributed randomly across conditions, and that the specific errors that non-native listeners made were on vowels that were disproportionately represented in high-density words. Suggestions for future research are presented.
Imai, S., Walley, A. C., & Flege, J. E. (2005) "Lexical frequency and neighborhood density effect on the recognition of native and Spanish-accented words by native English and Spanish listeners," Journal of the Acoustical Society of America, 117, 896-907.
Yoneyama, K. & Munson, B. (2010) "The influence of lexical factors on word recognition by native English speakers and Japanese speakers acquiring English: A First report," Journal of the Phonetic Society of Japan, 14-1, 35-47.
In this talk I’m going to present a brief overview of my PhD dissertation. In northern Italo-Romance dialects – with the exception of Ligurian as well as central and southern Venetian – -a is the only vowel that survived the general loss of all unstressed final vowels, which are supposed to have passed through a first stage of centralization in a pre-documentary phase, followed by vowel deletion. In the majority of these varieties the outcome of Proto-Romance -a is a low central vowel. However, there are some varieties in which this segment has undergone phonological processes: centralization, backing, fronting or — in particular — vowel harmony (VH).
The analysis of the systems of progressive VH in Swiss-Italian dialects is the focus of my research. This will then be extended to consider regressive VH phenomena and context-independent developments of the final vowel, in order to explore possible interactions of these phenomena with each other.
Though some of such VH systems have been mentioned in the literature, a systematic account is not yet available. In my research, I will consider the outcomes of -a in 40 dialects of the Swiss-Italian area where I carried out fieldwork; my dissertation aims to provide a detailed phonological description of the VH systems and some phonetic analyses of specific aspects.
In work on Articulatory Phonology (e.g., Browman & Goldstein, 1992) gestural representations of a word form were hypothesized to embody the information that is invariantly present in the signal, even when the form is produced in a variety of contexts that alter its acoustic and articulatory properties. This hypothesis was never really tested until recently. I will report on automatic speech recognition experiments (e.g., Mitra et al, accepted) that show that gestural information can be accurately recovered from the acoustic signal, and that augmenting standard acoustic representations with the extracted gestural information can lead to improvements in automatic word recognition, particularly in noise, where the improvements are dramatic. The gestural extraction algorithms are not based on statistics from real-word data, but rather artificial neural net learning from a purely theoretically-motivated model. The bottom line is that having a representation of the right kind is more important than getting the representation exactly right.
Browman, C. P., and Goldstein, L. (1992). "Articulatory Phonology: An Overview." Phonetica 49: 155–180.
Mitra, V., Nam, H., Espy-Wilson, C., Saltzman, E., and Goldstein, L. (Accepted pending minor revisions). "Tract variables for noise robust speech recognition." IEEE Trans. on Audio, Speech and Language Processing.
The establishment of a universal set of innate distinctive features is a hallmark of generative phonology. In recent years, the universality of distinctive features has been called into question from the viewpoint of both phonological theory (‘unnatural’ natural classes, ambivalent patterning, sign languages; e.g., Mielke 2008, Morén 2007) and language acquisition (Menn and Vihman 2010). I add to this debate through a study of the coronal phonology in languages with rich coronal systems. Such languages show different groupings of the coronals with respect to phonology, suggesting perhaps that a range of groupings are possible subject to physiological constraints rather than to innate features.
How do listeners accomplish the task of word segmentation given that in spoken language, there are no clear and obvious cues associated with word beginnings and ends? A given stretch of speech can be consistent with multiple lexical hypotheses, and these hypotheses can begin at different points in the input. In the French sequence l’abricot [labʁiko] ‘the apricot’, segmental information could be compatible with several competing hypotheses, such as l’abri [labʁi] ‘the shelter’, la brique [labʁik] ‘the brick’. Listeners are routinely confronted with such transient segmentation ambiguities, and in some cases ambiguities are total, as in Il m’a donné la fiche / l’affiche [ilmadonelafiʃ] ‘He gave me the sheet / the poster’. Yet the word recognition system is efficient, as listeners are rarely misled and generally segment correctly, retrieving the correct meaning.
In this talk, I am going to present a series of experiments examining the role of sub-phonemic acoustic-phonetic cues in speech segmentation and lexical access. We examined acoustic differences between phonemically identical sequences (e.g., l’affiche ‘the poster’, la fiche ‘the sheet’, both [lafiʃ]) and listeners’ discrimination and identification of these sequences. A series of off-line experiments (ABX & 2AFC paradigms) demonstrated that listeners can discriminate between and identify such utterances. Moreover the manipulation of the acoustic cues had an impact on the perceived segmentation: e.g., increasing the F0 in the /a/ of la fiche increased the percentage of vowel-initial (affiche) responses. A series of on-line experiments (cross-modal identity and fragment priming) suggested that listeners retrieve on-line the correct segmentation and modulate activation of targets and competitors in favor of the correct candidate, without ruling out alternative interpretations. These results provide further evidence for fine-grained lexical access and suggest that the recognition system exploits sub-phonemic information to guide segmentation towards the beginning of words. The implications of these findings for models of word recognition will be discussed.
Articulatory plans are learned and modified via information from multiple internal and external sources: auditory, somatosensory, visual, social, etc. This experiment measures how flexible those plans are, and whether that flexibility is tied to high-level properties of a speaker's native language or individual production habits.
During the experiment, specialized software altered speakers' formants in real time. Speakers produced CVC words but heard the vowels in those words with a higher F2. Previous studies have shown that speakers compensate for the mismatch between heard and expected auditory feedback by opposing the alteration: for example, a talker would compensate for an F1 increase in /ɛ/ by producing a vowel with a lower F1. Here, subjects heard their F2 altered in several vowels with a range of somatosensory and acoustic properties. On the whole, subjects compensated more for vowels with few phonological neighbors, providing some evidence for an influence of top-down knowledge about phonological categories on the processing of incoming auditory feedback. However, there was substantial individual variation in vowel production and compensation across subjects.
A second analysis asked whether a subject's formant trajectory during experiments altering vowels in several regions of vowel space could be predicted by the familiarity or phonological identity of acoustically similar vowels. Two models were tested -- one in which the subject's vowel space was determined from his casual speech, and one in which the subject's vowel space was determined from vowels in citation form words. Several regression and optimization analyses showed that the citation form vowel space matched the formant trajectory more consistently than the casual speech vowel space.
This analysis suggests that articulatory plans are modified with respect to high-level phonological categories and information about the speaking situation.
In this talk I will report some results from my dissertation research that examines how speakers of American English produce, perceive, and repeat the high back vowel /u/ in fronting and non-fronting contexts. Results from production study show that acoustic distance between fronted /u/ and canonical /u/ remains approximately constant across the changes in speech rate, suggesting that the fronted variants are produced by controlled articulation. The results also show that fronted /u/ has much smaller variability than canonical /u/. This variability in production correlates with the variability in category judgment in perception. Perception data show considerable individual variation in perceived /i/-/u/ category boundaries both in fronting and non-fronting contexts. Related to this variation, results from vowel repetition study show that ambiguous vowel stimuli were repeated as categorically different sounds across subjects depending on how they interpret each stimulus. Based on these observations, I argue for the following propositions: 1) controlled coarticulations provide precondition of phonologization by ensuring distinct phonetic distribution of exemplars; 2) speech perception is influenced by phonological knowledge that includes knowledge of the distribution of within-phoneme variants; and 3) it is the listener’s interpretation of acoustic patterns that gives rise to innovative pronunciation.
At this talk, members of the 2010-2011 Field Methods class will present their most recent findings on Abo, a Bantu language spoken in Cameroon.
At this event, scheduled outside of Phorum's normal time, the third-year cohort will present 20 minute presentations on the findings of their Qualifying Papers. This event will take place from 3:00 to 5:45 in 370 Dwinelle (G-Level).
In this talk I will introduce a new theoretical concept: information utility. I will discuss the need for a different information-theoretic approach in linguistics, and demonstrate how the new approach can be used to address Weinreich, Labov, and Herzog’s (1968) actuation problem. The actuation problem revolves around the reasons that lead some language to choose one structural change and other languages to choose other structural changes. In both UK and American English, for instance, there are several /t/-weakening processes, a rather uncommon process cross-linguistically. Likewise, in many different dialects of Arabic /q/ is weakened. I show that these parallel processes are not due to language contact, and that the key to solving the actuation problem for weakening processes lies in balancing information utility with effort. While articulatory effort is relatively constant cross-linguistically, the information utility of segments emerges from their language-specific distibutions, and is therefore different in each language. In order to demonstrate how utility and effort interact, I present an OT-based model in which faithfulness constraints instantiate information utility, and markedness constraints represent effort. With this reinterpretation of faithfulness and markedness, I show that the proposed model predicts the observed distribution of weakening processes.