1 / 1

How General is Lexically-Driven Perceptual Learning of Phonetic Identity?

How General is Lexically-Driven Perceptual Learning of Phonetic Identity?. Tanya Kraljic and Arthur G. Samuel. Experiment 2: Background Expt 2 differs from Expt 1 in two ways:

derick
Download Presentation

How General is Lexically-Driven Perceptual Learning of Phonetic Identity?

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. How General is Lexically-Driven Perceptual Learning of Phonetic Identity? Tanya Kraljic and Arthur G. Samuel Experiment 2: Background Expt 2 differs from Expt 1 in two ways: 1. Phonemes at exposure are different: We use /s/ and /S/ (SH), rather than /d/ and /t/. Rationale: Using fricatives instead of stops enables us to obtain a much larger perceptual learning effect, and therefore to look at the extent to which the learning effect goes away under various unlearning conditions. 2. A 20-minute intervening ‘Unlearning’ phase is added between Exposure and Test(the task is card sorting: subjects hear descriptions of 20 cards; the descriptions are in the Same or Different voice as exposure, and they either do or do not have normal /s/ or /S/ in them. The task is described in more detail at the end of the poster):Rationale:Given that exposure to some odd /s/ or /S/ has caused changes in speech categorization, this unlearning phase addresses whether the representations revert to their prior settings, and if so, what causes them to do so. There are three possibilities that we test. The system may: a. revert to prior settings simply after time has passedb. require further exposure to normal /s/ and /S/, REGARDLESS OF SPEAKER, to revert c. require further exposure to normal /s/ and /S/ by the SAME SPEAKER, in order to revert Experiment 1: Method Phase I - Exposure (Task: Lexical Decision): 2 X 2 Between subject Voice at Exposure General Conclusions Taken together, the results from Expts 1 and 2 have several implications for the questions we began with: We have found 2 kinds of evidence that perceptual learning generalizes across speakers: Expt 1: the basic /d/-/t/ effect transferred to a new speaker Expt 2: unlearning was brought about by normal /s/ or /S/ from the same speaker, and also from a new speaker However, we also have 2 kinds of evidence suggesting some level of speaker specificity: Expt 2: the CD condition suggests some remaining perceptual learning when the speaker is new Ongoing projects: where we use fricatives rather than stop consonants suggest a speaker specific component. Expt 1: For stop consonants (/d/ and /t/), perceptual learning appears to operate at a featural level, rather than a phonemic one. Expt 2: The representations revert back to their original form when additional input corrects the odd phonemes from exposure. Simply allowing time to pass did not cause the representations to revert. Introduction Speech sounds (phonemes) are realized differently across different speakers - this variation is due to dialectal differences, foreign accent, speech rate, and other factors. But, representations of these phonemes have typically been thought to be stable abstractions from these many instantiations. Perceptual Learning: Listeners use their lexical knowledge to actually adjust their phonemic representations in a way that is consistent with the pronunciation they are hearing (Norris, McQueen & Cutler, 2003) Male 20 words where /d/ is replaced with /?/ (secon?ary)20 /t/-intact words (cafeteria)60 filler words + 100 filler nonwordsno /d/, /t/, /b/, or /p/ 20 /d/-intact words (secondary)20 words where /t/ is replaced with /?/ (cafe?eria)60 filler words + 100 filler nonwords no /d/, /t/, /b/, or /p/ Female 20 words where /d/ is replaced with /?/ (secon?ary)v20 /t/-intact words (cafeteria)60 filler words + 100 filler nonwords no /d/, /t/, /b/, or /p/ 20 /d/-intact words (secondary) 20 words where /t/ is replaced with /?/ (cafe?eria) 60 filler words + 100 filler nonwords no /d/, /t/, /b/, or /p/ 1.How general is perceptual learning with respect to speaker? ?D Phoneme Exposure Conditions ?T (Neutral condition: N) Basic Perceptual Learning For example, an average labeling function for a continuum of seven sounds that range from more /d/-like to more /t/-like (with the midpoint (/?/) being ambiguous between the two) generally looks like this: But, this function actually shifts depending on people’s experience with the sounds on the continuum, and where they occur in words. (Corrected phonemes, Different Speaker: CD) (Corrected phonemes, Same speaker: CS) Phase II - Test (Task: Category ID)2 X 2 Within Subjects Identify all test items: SameVoice and Different Voice as exposure than exposure /d/-/t/ and /ada/ - /ata/ /ada/ - /ata/ /IdI/ - /ItI/ /IdI/ - /ItI/ /b/-/p/ continua /aba/ - /apa/ /aba/ - /apa/ /IbI/ - /IpI/ /IbI/ - /IpI/ This within-subjects design allows us to test generalization to new speakers (Same Voice vs. Different Voice) and to new phonemes (Same DT Phonemes vs. New BP Phonemes). Experiment 2: Method Phase I - Exposure (Lexical Decision) same as Experiment 12 conditions, between subjects: ?S Exposure ?SH Exposure (Male speaker) (Male speaker) Phase Intervening (20 minutes) Card Sorting Task3 conditions, between subjects: Neutralnormal /s/ normal /s/, Neutral normal /S/ normal /S/ (Time) Female spkr Same spkr (Time) Female spkr Same spkr Phase II - Test (Category ID)same as Experiment 1 /asi/ - /aSi/Original (male) Speaker 2.How general is perceptual learning with respect to speaker? CS N N 3. What happens after the representations have been adjusted? CS CD CD /d/ /t/ /?/ Experiment 1: Results a. Perceptual learning generalizes to a new speaker (reliable training effect, no interaction with speaker): * *Note: Although the effect here is reliable across speaker, ongoing work in our lab suggests that the answer may be different for fricatives(/s/ and /z/, /s/ and / S/). If people are exposed to the /?/ sound in words in which a /d/ normally occurs (e.g., croco/?/dile, secon/?/ary), they label more items on the /d/-/t/ continuum as /d/: However, if the /?/ sound occurs in words where a /t/ should be (e.g., cafe/?/eria, mili/?/ary), people label more of the items as /t/: That is, hearing an ambiguous version of a phoneme expanded that phoneme’s category, to include more of such ambiguous tokens. Experiment 2: Results People who received the Neutral unlearning condition (20 minutes intervening between exposure and test, with no intervening /s/ or /S/) showed large perceptual learning effects (p=.01). When the intervening condition contained good /s/ or /S/ phonemes from the Same speaker that they were originally exposed to (with ‘bad’ /s/ or /S/), there was no perceptual learning (p=.585). Same Voice(Male at exposure/Male test OR Female at exposure/Female test) Different Voice(Male at exposure/Female test OR Female at exposure/Male test) N CS /d/ /?/ /t/ References: Norris, D., McQueen, J., & Cutler, A. (2003). Perceptual learning in speech. Cognitive Psychology, 47, 204-238. Neutral Unlearning Condition(conceptually equivalent to Time passage) CS Unlearning Condition(Good /s/ or /S/ phonemes, Same speaker) N CS Many, many thanks to Donna Kat for all her help on these projects! /t/ (or /p/) /d/ (or /b/) /t/ (or /p/) /d/ (or /b/) b. Perceptual learning also generalizes to new phonemes (/b/-/p/) that share the same featural relationship as the phonemes listeners were exposed to (/d/-/t/) (reliable training effect, no interaction with phoneme): More information on Experiment 2’s Intervening Task: Subjects received verbal descriptions - presented over headphones - to sort a deck of 20 cards with tangrams (random shapes) on them. Examples of tangrams: We used tangrams to discourage any implicit naming, since our critical manipulation is the phonemes subjects hear (and do not hear) in the descriptions for each card. Subjects sorted 4-5 decks, until 20 minutes had passed. Critically, subjects were assigned to one of three conditions:Neutral (N)-descriptions did not contain any /s/ or /S/ (E.g., an orange letter “Y”) Corrected - descriptions included either good /s/ or good /S/ (i.e., whichever was the normal version of the /?/ phoneme they had been exposed to in Phase I)Same Voice (CS):Same male Voice as in Phase IDifferent Voice (CD): New (female) Voice (E.g., Phase I ?S: An orange letter “V” sitting on a square. A Phase I ?SH:An orange one shaped like the letter “Y” /d/ /?/ /t/ Same Phoneme(D or T exposure/DT test continuum) Different Phoneme(D or T exposure/BP test continuum) Our Questions (e.g., learning a particular person’s accent vs. applying this changed phonemic‘category’ more generally) (e.g., adjusting the category for a particular phoneme vs. making a more general, feature level adjustment) Do they return over time to the original form, or do they remain in the adjusted form until some further input causes another change? /s/ /s/ /S/ /S/ Experiment 1:How general is perceptual learning? With respect to: When the intervening condition contained good /s/ or /S/ phonemes from a Different speaker than the exposure and test speakers,most of the perceptual learning effect was gone (p=.174). CD a.speaker CD Unlearning Condition(Good /s/ or /S/, Different speaker) CD b. phoneme /d/ /p/ /t/ /b/ Experiment 2:What happens after the representations have been adjusted? Experiment 1: Conclusion Perceptual learning appears to generalize both to new speakers and new phonemes. /s/ /S/

More Related