Edinburgh Disfluency Group SCHOOL of PHILOSOPHY, PSYCHOLOGY and LANGUAGE SCIENCES UNIVERSITY of EDINBURGH 27 March 2009 Speech Bacground Experimens Findings end o affec: words Bonnie Prince Billy Prince Bonnie Billy syllables Capain Beefhear Beefain Caphear morphemes Beasie Boys Beasies Boy phonemes Johnny Cash Connie Jash esimaed incidence 31 160 per 100K words abou as frequen as saying London or war (Garnham e al., 1981; Shallice & Buerworh, 1977) informaive abou he building blocs of speech (Fromin, 1971; Garre, 1975) Speech The Edinburgh Disfluency Group Speech Bacground Experimens Findings Phoneme Subsiuion Errors researchers ineresed in imperfec speech wha can errors ell us abou he way speech is produced? how do he imperfecions in speech affec communicaion? wha causes difficulies in speaing? approaches include experimenal wor compuaional modelling surveys, quesionnaires, oher invesigaions members largely from Edinburgh and Queen Margare Universiies collaboraors in Belgium, he US... Which of us has no fel in his hear a half-warmed fish? collecions of phonemic errors dae bac o 19h C (e.g., Meringer & Mayer, 1895) many are noeboo collecions and may no be represenaive (e.g., Fromin, 1971; Vousden e al., 2000) hp://edgwii.wiido.com Today Speech Speech Bacground Experimens Findings Measuring 1 Bacground Experimens Findings 2 Models of Speech Feedbac The Monior noeboo collecions only record wha people heard people end o hear phonemic speech sounds as belonging o one caegory or anoher (Liberman, 1997) need o measure wha people say naural error incidence is very low need o encourage error producion big fun
Speech Bacground Experimens Findings Eliciing Errors: The WOC Tas Speech Bacground Experimens Findings Measuring Wha People Say simple as designed o elici speech errors wors by creaing compeiion wih correc uerance correc response jed pul urve eam display jed pul eam urve Speech Bacground Experimens Findings Elecropalaography (EPG) demonsraion of he WOC (Word Order Compeiion) Tas allows measuremen of he ongue s conac wih he palae (e.g., Edwards e al., 1997; Scobbie e al., 2004) Speech Bacground Experimens Findings WOC Experimen Speech Bacground Experimens Findings Elecropalaography (EPG) 47 paricipans produced 96 experimenal ( ) iems each speech was ranscribed and coded for 7 paricipans we also oo ariculaory recordings (using EPG) experimenal iems always began wih,d,g barn door
Sample EPG Daa Speech Bacground Experimens Findings Speech Bacground Experimens Findings Quanifying Differences in EPG ave gub ave gub Iniial Findings Speech Bacground Experimens Findings Speech Bacground Experimens Findings Comparing EPG recordings Table: Percenages of errors recorded in he experimen Audiory EPG 47 pps 7 pps 7 pps subsiuions 1.1% 3.0% 3.0% double ariculaions 2.2% 13.2% oher 17.6% 18.0% couning double ariculaions alone, over 80% of relevan errors are no heard by ranscribers all of he informaion we have abou errors excludes mos slips of he ongue! Dimension 2 4 2 0 2 4 analysis gives a disance beween any wo recordings lie a mileage able on a map can recreae a map using mulidimensional scaling 5 0 5 10 Dimension 1 16 s compared o 16 s from conrol ( ) condiion Speech Bacground Experimens Findings Wha is an error? Speech Bacground Experimens Findings Redefining Speech Error insead of rying o wor ou wha is an error, measure he difference beween condiions which of hese are good s and which are errors? he classificaion problem isn solved by EPG need a way o compare ariculaions o each oher when you say in ave gub g in ave gub?, how close is i o
Resuls Speech Bacground Experimens Findings Speech Feedbac Monioring More on Our Error Research when you say in ave gub ave gub? Ariculaory Variaion ( ) 0 1 2 3 4 5 nonlexical, how close is i o g in * Compeior closer when maing an error would have resuled in a word (ave gub gave ub ) han no (afe gup gafe up ) lexical g - + A V voicing place ariculaion WORDS release vowel onse "voiceless" "voiced" Implicaions Speech Feedbac Monioring Speech Feedbac Monioring Self-Monioring our findings confirm a lexical bias in speech errors words exer an influence on ongue movemens someimes hose ongue movemens are influenced so much ha liseners hear differen words bu errors are in he ears of he beholder ca flaps ool is ofen we are able o edi our planned speech may provide an alernaive explanaion of lexical bias more liely o edi hings ha clearly aren words acually a complicaed sory (Harsuier e al., 2005) Speech Feedbac Monioring The Feedbac Model Speech Feedbac Monioring The Monioring Loop GAVE WORDS concepualizer message generaion monioring discourse model siuaion nowledge encyclopedia ec. preverbal message "parsed speech" g ave ub ONSETS RIMES PHONEMES formulaor grammaical encoding surface srucure phonological encoding lexicon lemmas forms speech comprehension sysem phonemes feed bac acivaion o words compeior ( ) phonemes become misacivaed words feed forward o compeior phonemes if here are no relevan words, acivaion isn increased phoneic plan (inernal speech) ariculaor over speech phoneic sring audiion (Level, 1989)
Disfluency in Speech Speech Feedbac Monioring Thans Speech Feedbac Monioring deciding wha o say aes effor and ime so does monioring edis may be noiceable in speech oupu disfluencies in speech are acually aemps o eep hings fluen fillers (um, er) repeiions (over) repairs nex wo als abou disfluency as opposed o errors much of he wor discussed in his al in collaboraion wih Corey McMillan (Universiy of Pennsylvania) also Robin Licley, Rob Harsuier, Suzy Moa, and ohers hp://edgwii.wiido.com/ funding: NIH/NIDCD, ESRC Conclusions Speech Feedbac Monioring here s no such hing as a speech error when he plan changes enough, speech may be misheard ofen here is no discernable error even hough he ongue is inaccurae hese inaccuracies ell us abou he srucure of speech in paricular, i s a highly ineracive sysem in which differen levels all affec each oher people monior heir own speech plans for problems; someimes rying o fix hese resuls in disfluency